WorksComputer Science
Prompt Re-Specification and Target-Mask Availability in Mask-Conditioned Visual Reinforcement Learning for Moving-Platform Landing
Abstract
Visual reinforcement learning (RL) policies that land a drone on a moving platform can fail when the landing pad looks different from the one used in training. Using binary target masks as the visual input removes most pixel-level appearance information while preserving the spatial representation of the target, and lets a text prompt specify the target. In simulation, we test whether a frozen (fixed-weight) mask-based DrQ-v2 policy can handle a new pad when only the prompt of its fine-tuned SAM 3 detector changes. A banana-themed pad lowered landing success from 75.6% to 12.0%, and SAM 3 rarely detected it under the original prompt. A revised prompt, chosen after testing several candidates, restored success to 68.4% with no weight updates. In a separate experiment with a policy trained on ground-truth masks, filling in masks that were missing while the pad was visible helped landing far more than the tested mask-centroid and area corrections. Tracking the pad over time supplied masks more often and also kept producing them after the pad was confirmed absent from the rendered image, but it did not establish a reliable improvement in landing success. Whether a usable mask reaches the policy at all thus appears to be an important limiting factor for this approach.
Introduction
Autonomous landing on moving platforms is an important capability for unmanned aerial vehicles (UAVs) in mobile recharging, logistics, inspection, and maritime operations [1]. Cameras offer lightweight sensing without requiring communication with the platform or external positioning infrastructure [2, 4], and reinforcement learning (RL) can learn tracking, alignment, and descent together in one closed-loop policy. Ladosz et al. [4] trained such a policy with DrQ-v2 [10]. It maps camera images and onboard velocity and attitude to velocity commands, without access to the position or velocity of the platform.
A policy that learns from unsegmented camera images can become tied to the appearance of the landing pad it saw during training [4, 9]. If the pad is repainted or replaced, the policy may need new data and retraining. Promptable segmentation models [11, 12] offer another route. A perception model finds the target from a text prompt and passes it to the controller as a binary mask. Most of the appearance of the pad then never reaches the policy, and a changed pad can be described by a new prompt. This alone does not make landing robust to appearance, because the perception model still has to detect the new pad. When direct detection fails, the perception frontend falls back on temporal propagation, as in video segmentation models with memory [15], and this propagation can also fail or carry outdated target information into later frames [16]. In closed loop, the frames that the policy sees also depend on its own earlier actions, and frame-level segmentation accuracy need not predict downstream performance [17].
This paper asks whether a frozen mask-conditioned RL landing policy can adapt to a new landing-pad appearance when only the segmentation prompt is changed, without updating any network weight, and what limits this adaptation. Throughout the paper, “frozen” means that the weights of a network are not updated after its training, so a frozen network has the same weights in every condition in which it is evaluated. We study the question in simulation with a DrQ-v2 actor that observes masks from a SAM 3 detector fine-tuned for the task. When a banana-themed pad replaced the training pad, direct detections and landing success collapsed under the original prompt. A revised prompt brought back most of the detections and most of the landing success. A substitution diagnostic with a separately trained ground-truth-mask policy points to the property involved. With raw SAM-derived masks, this policy almost never landed. Filling in only the masks that were missing while the target was visible restored landing, whereas correcting the centroid and area of the masks that were produced did not. The perception frontend already uses temporal tracking to supply masks through detection gaps. Reinitialising the tracker at each direct detection made the target available more often, but it also produced more nonempty masks after render-confirmed target absence, and the resulting difference in landing success was not established.
The central contribution of this work is an empirical account of prompt-conditioned segmentation as the observation interface of a frozen RL landing policy when the target appearance changes. It identifies target-mask availability as a key limitation for control. Three connected investigations support it. A matched appearance and prompt evaluation (Section IV) shows that revising the prompt recovered most of the landing performance lost under the tested appearance change. A frozen-policy substitution diagnostic (Section V) shows that, among the interventions tested, restoring missing masks mattered far more for landing than correcting mask geometry. A comparison of temporal memory management (Section VI) shows that greater target continuity comes with more nonempty output after renderer-confirmed target absence and does not automatically improve landing. The results concern one substantially different pad appearance, a prompt revised by the authors, and one trained actor, all in simulation.
Related Work
Moving-Platform Landing
Classical systems treat moving-platform landing as modular estimation and control. Image-based visual servoing uses image-space target features and optical flow [6], [7], and feed-forward variants add an estimate of the platform velocity [5]. Falanga et al. [2] predict the platform motion with an extended Kalman filter while detections are unavailable, and Paris et al. [3] plan trajectories whose final velocity matches the estimated platform motion. These systems bridge visual gaps with an explicit state estimate, but they rely on an engineered marker and a structured pipeline. Learning-based methods replace part or all of this pipeline. They use extracted relative state [8], sub-task decomposition [1], or image-based policies transferred with domain randomisation [4], [9]. The end-to-end policy of Ladosz et al. [4] assumes that the pad design was seen during training, and its environment and learner are the closest to ours. Other recent methods handle partial observability inside the policy, with a long short-term memory network over landing-specific keypoints [13] or recurrent RL over noisy and missing state measurements [14]. We keep the actor feed-forward and deal with appearance and missing observations in the perception interface.
Perception Representations for Learned Control
The representation passed from perception to a learned controller decides which errors the controller has to tolerate. A driving policy shielded from raw perceptual input by a modular perception layer transferred from simulation to a real vehicle [24], and agents given intermediate vision representations learned faster and generalised better than agents learning from pixels alone [25]. In visual RL, images masked with Segment Anything improved generalisation to unseen visual conditions [26]. Segment Anything [11] segments objects from visual prompts, and SAM 3 [12] extends promptable segmentation to concepts given as text, so a text prompt can describe a new target without retraining the controller. In our setting the appearance change concerns the landing target itself, and the adaptation is made through the text prompt. Bounding boxes and keypoints are more compact alternatives to masks; we do not compare interface types. Because perception metrics need not reflect downstream performance [17], we evaluate the interface by closed-loop landing.
Temporal Memory in Video Segmentation
Video object segmentation carries information across frames. SAM 2 [15] conditions each frame on a streaming memory of earlier predictions, XMem [19] organises memory into sensory, working, and long-term stores, and DAM4SAM [22] adds distractor-aware memory selection to SAM 2.1. Memory can also spread mistakes: SAM2Long [16] shows that the greedy memory selection of SAM 2 lets a wrong or missed mask affect later frames. These methods are evaluated by segmentation accuracy on recorded video. We instead measure how memory management changes the observations that a closed-loop controller receives, and how landing outcomes differ.
Task and System
Landing Task
A quadrotor lands on a non-communicating moving platform using only onboard visual and ego-state observations. The simulation environment is built on gym-pybullet-drones [20]. It is adapted from the moving-platform landing environment of Ladosz et al. [4] and shares its nominal start height, horizontal reset range, and maximum platform speed. The vehicle is the Crazyflie 2.X model of the simulator with a rigid, downward-facing camera, and a proportional–integral–derivative (PID) controller tracks its velocity commands. The landing target is a 1 m × 1 m pad bearing a fixed greyscale circular landing-marker texture, which is not randomised during training.
Each episode starts with the UAV at a height of 10 m and a horizontal offset from the platform drawn uniformly from m along each axis. The platform moves along a straight line at a speed drawn uniformly from m s and a heading drawn uniformly from . The physics simulation runs at 240 Hz, the PID controller at 24 Hz, and the actor at 12 Hz. Fresh camera images are segmented at 6 Hz, so the actor acts on each mask for two consecutive decisions. An episode ends at the first contact or after 30 s. A landing succeeds when the UAV touches down with its horizontal collision footprint, of radius 0.06 m, inside the pad. Both horizontal offsets from the pad centre, measured in the pad frame, must therefore be at most 0.44 m, and no limit is placed on touchdown velocity. Any other contact is a crash, and an episode still airborne at 30 s is a timeout. Table 1 summarises the parameters.
| Parameter | Value |
| Physics / PID rate | 240 Hz / 24 Hz |
| Actor rate | 12 Hz |
| Perception update rate | 6 Hz |
| Camera model | pinhole, no distortion |
| Focal length | 63.455 px |
| Principal point | 41.5 px |
| Field of view | approx. |
| Mask resolution | |
| Camera near plane | 0.0397 m |
| Landing-pad size | |
| Initial height | 10 m |
| Initial horizontal offset | per axis |
| Platform speed | |
| Platform heading | |
| Episode horizon | 30 s (360 actor decisions) |
Table 1. Landing-task parameters.
Perception Frontend
The perception frontend turns each fresh camera image into the mask and perception metadata that the actor observes. It uses three frozen models. The direct detector is a SAM 3 image segmenter [12] that was fine-tuned for this project with the text query “landing pad” and then frozen (Appendix D describes what is known about this fine-tuning). At inference it receives the same prompt and uses confidence and mask thresholds of 0.5 at an internal resolution of 1008 pixels. DAM4SAM [22], a tracker built on SAM 2.1 Hiera-Tiny [15], propagates the target when direct detection fails, with its native object-presence processing enabled. A CLIP ViT-B/32 image encoder [21] supplies similarity metadata.
A direct SAM 3 mask always takes precedence. Let be the direct mask at perception step and the mask that the tracking session propagates to the current frame. The mask passed to the actor is
In the base frontend, with which the actor was trained, a tracking session is initialised on the first nonempty direct mask of an episode and then advanced on every fresh frame. Later direct detections override the output for their own frame but do not reinitialise the session, so after a gap the tracker continues from its own accumulated memory. The base frontend therefore already has temporal memory and is not a detector-only baseline. Section VI introduces a frontend that manages this memory differently.
On an NVIDIA L40S server GPU, one perception update took 158.2 ms on average with the base frontend and 148.5 ms with Temporal Recovery (Section VI), against the 166.7 ms available at 6 Hz. Onboard latency was not measured.
Observation, Action, and Policy
Figure 1 summarises the pipeline. The actor never receives the pose, velocity, or heading of the platform. Its observation consists of the three most recent binary masks,

Figure 1. Overview of the shared perception, policy, and control pipeline. Direct SAM 3 masks take precedence and temporal predictions provide fallback observations. The evaluated frontends differ in their temporal anchoring (Section VI). CLIP similarity is part of the perception metadata and does not reject masks.
and a low-dimensional history with one row per frame. Each row holds the world-frame velocity of the UAV (3 entries) and its attitude quaternion (4), which stand in for onboard measurements. It also holds eight perception metadata entries, computed only from the images and the perception outputs. These are one-hot flags marking a direct, propagated, or empty mask (3), the age of the last direct detection (1), the native segmentation quality score and its availability (2), and a CLIP similarity between the current mask crop and templates from earlier direct detections, with its availability (2). The flags and the detection age let the actor tell direct, propagated, and missing observations apart.
The actor outputs a world-frame velocity command , limited to horizontally and vertically.
The policy is trained with DrQ-v2 [10]. The actor encodes the mask stack with the DrQ-v2 convolutional encoder, concatenates the flattened history, and maps the result through a 50-unit linear layer with BatchNorm1d and tanh and two 512-unit hidden layers to a bounded velocity command. Training uses an asymmetric actor–critic [18]. Only the critic also receives the position and velocity of the platform relative to the UAV. This privileged state can ease value estimation, but it never enters the actor and is absent at inference. The reward combines terminal rewards of +140 for success, −50 for a crash, and −45 for a timeout with a small time cost and potential-based shaping [23] on the horizontal distance, the horizontal relative speed, and the relative altitude, all computed from simulator state. The actor was trained once, with one seed, for 1.25 million decisions, first under an easy-to-hard curriculum [1], [13] and then under the nominal and a speed-prioritised reset distribution (Appendix § 3.4). Every comparison in Sections IV and VI uses this frozen actor with exploration noise disabled.
| Speed () | A | B | C | D |
| 0.00 | 35/6/9 (70%) | 4/20/26 (8%) | 34/8/8 (68%) | 31/6/13 (62%) |
| 0.25 | 38/2/10 (76%) | 6/13/31 (12%) | 34/8/8 (68%) | 42/3/5 (84%) |
| 0.50 | 42/3/5 (84%) | 8/6/36 (16%) | 40/7/3 (80%) | 44/2/4 (88%) |
| 0.75 | 42/1/7 (84%) | 6/4/40 (12%) | 37/6/7 (74%) | 41/3/6 (82%) |
| 1.00 | 32/1/17 (64%) | 6/2/42 (12%) | 26/4/20 (52%) | 36/0/14 (72%) |
| All | 189/13/48 (75.6%) | 30/45/175 (12.0%) | 171/33/46 (68.4%) | 194/14/42 (77.6%) |
Table 2. Appearance and prompt evaluation with the frozen actor and the Temporal Recovery frontend, 50 paired episodes per speed and condition. Entries are success/crash/timeout (success rate). A: original pad, original prompt; B: banana-themed pad, original prompt; C: banana-themed pad, re-specified prompt; D: sampled within-family greyscale pad texture, original prompt.
Evaluation
The closed-loop conditions are the nominal task distribution (P), an exactly stationary platform (S0), and platform speeds restricted to [0, 0.25), [0.25, 0.5), [0.5, 0.75), and [0.75, 1] (S1–S4). S0 is included because the nominal sampler produces an exactly stationary platform with probability zero. In every comparison, the evaluated configurations face the same episode seeds and reset states, so outcomes are matched episode by episode. We report the success rate (SR) together with crash and timeout counts, and compare matched success with exact two-sided McNemar tests.
Two perception metrics describe what the actor receives. Terminal usable-target coverage is the fraction of renderer-visible fresh camera exposures in the final five simulated seconds of an episode for which the returned mask is usable. These exposures are selected using their recorded exposure-associated simulator step counters. A mask is usable if it overlaps the rendered target by at least 20 pixels, with an intersection over union (IoU) of at least 0.30 and a recall of at least 0.50, and an empty output on a visible target counts as unusable. The post-absence output rate is the fraction of fresh frames after the last renderer-visible target frame of an episode on which the frontend still returns a nonempty mask. Each configuration generates its own trajectories, so both metrics are descriptive.
Appearance Change and Prompt Re-Specification
The appearance evaluation changes what the perception model has to recognise while the actor, the perception network, and the pad geometry stay fixed. It uses the frozen actor with the Temporal Recovery frontend, which reinitialises the tracker at each direct detection (Section VI). Conditions A, B, and C share seeds and reset states at each of five fixed platform speeds (0, 0.25, 0.5, 0.75, and 1 ), with 50 matched episodes per speed. Comparing A with B isolates the appearance change under the original prompt “landing pad”, and comparing B with C isolates the prompt change on the same appearance. The revised prompt, “banana-printed landing pad”, is supplied only to SAM 3; CLIP receives no text, although its image templates come from the direct detections. The weights of the actor and of every perception network are identical in all conditions. Condition D uses the seeds and reset states of A but gives each of its 250 episodes a different pad texture. The textures are sampled without replacement from a 424-image greyscale library of variations within the landing-marker family of the nominal pad.
The revised prompt came from exploratory testing. Several candidate prompts were compared on fixed images of the banana-themed pad, without running landing episodes and on samples separate from the evaluation episodes. The evaluation episodes themselves had already been run once with a memory-restricted variant of the frontend (Appendix C) before the evaluation reported here. The result therefore shows prompt-mediated recovery in the tested setting, not prospectively validated generalisation. Appendix § 3.4 gives the selection details.
tbl:II reports the outcomes. Under the original prompt, the banana-themed pad lowered success from 75.6% to 12.0%, and most of the failures (175 of 220) were timeouts. Direct detections fell from 63.3% to 8.8% of perception frames. The remaining frames shifted to propagated masks, which rose from 21.3% to 50.2%, and to empty observations, which rose from 15.4% to 41.0%. Whole-episode usable-target coverage dropped from 92.8% to 74.3%. Changing only the text prompt brought direct detections back to 61.8% of frames, coverage to 89.1%, and success to 68.4%. Both A and C exceeded B at all five speeds. Pooled over the five speeds, 163 episodes succeeded only in A and 4 only in B (exact two-sided McNemar test, ), and 7 succeeded only in B and 148 only in C ().
The recovery was incomplete. Success with the revised prompt stayed below that on the original pad by 7.2 percentage points overall and by 2 to 12 points at each speed (57 A-only against 39 C-only successes, ). Condition C also had 33 crashes against 13 in A, with a similar number of timeouts (46 against 48). No degradation was detected with within-family textures. Success was 77.6% in D against 75.6% in A (35 A-only against 40 D-only successes, ), per-speed differences ranged from −8 to +8 points, and the perception statistics were similar (direct 63.5%, propagated 21.4%, empty 15.0%, coverage 93.1%).
The binary mask keeps raw appearance out of the actor input, but the perception model still has to detect the target. Landing success fell and recovered together with the availability of direct detections, which points to target-mask availability as an important limiting factor in this setting. This is an association rather than an isolated effect. The perception fractions come from different closed-loop trajectories, and the experiment does not separate availability from other changes in the observations, such as mask quality, metadata, temporal history, and the resulting trajectories. The next section examines availability with a controlled substitution.
| Observation condition | Screen (50 ep.) | Confirmation (200 ep.) |
| Ground-truth mask | 41/50 | 165 (82.5%) |
| Raw SAM-derived mask | 0/50 | 3 (1.5%) |
| Centroid, area, or joint correction | 0/50 | – |
| Visible-dropout oracle repair | 42/50 | 174 (87.0%) |
Table 3. Frozen-policy substitution diagnostic with a separately trained ground-truth-mask actor, the same fine-tuned SAM 3 checkpoint as the main study, and the earlier protocol of Appendix A. Each repair modifies one class of SAM error; the actor is unchanged. These results are not comparable with those of the main actor.
Target-Mask Availability Diagnosis
The appearance evaluation does not intervene on the masks themselves. To identify which segmentation errors matter for a mask-conditioned controller, we replace the observations of a frozen policy one error class at a time. This requires a policy that can also act on renderer-derived ground-truth masks. The diagnostic therefore uses a separate DrQ-v2 actor trained on such masks under an earlier and simpler protocol, with shorter episodes, lower starts, slower platforms, a smaller pad, and a different reward and success criterion (Appendix A). The raw-mask condition uses the same fine-tuned SAM 3 checkpoint and prompt as the main study, without any temporal fallback. The diagnostic therefore differs from the main study in the actor, the observation interface, and the protocol. Its success rates cannot be compared with those of the main actor, and its ground-truth condition is not an upper bound for that actor. What it can show is which class of SAM error changes landing when the controller is held fixed.
tbl:III gives the results. In a 50-episode screen, the ground-truth condition succeeded in 41 episodes and raw SAM-derived masks in none. Correcting the centroid, the area, or both in every nonempty SAM mask also gave 0 of 50. An oracle repair that replaces an empty SAM output by the ground-truth mask only when the target is visible in the rendered image, leaving every nonempty prediction unchanged, gave 42 of 50. Only this repair was carried forward to a 200-episode confirmation. There, success was 82.5% with ground-truth masks, 1.5% with raw SAM-derived masks, and 87.0% with the repair.
The nonempty SAM masks were geometrically accurate. On the diagnostic calibration set they had a mean IoU of 0.898, a mean centroid error of 0.70 pixels, and a mean area bias of −7.9%. Their availability was the weak point. Recall was 96.5% when the target was fully visible, 46.7% when it was partially visible or at the image border, and 71.7% over all visible frames. Visible targets also tended to be missed in bursts, with a median length of five perception frames. For the interventions tested, correcting missing masks was much more consequential for landing than correcting the tested geometric errors. The diagnostic covers only centroid and area corrections and does not show that mask geometry is irrelevant in general.
The main actor shows the same weak point in closed loop. It was trained on SAM-derived masks, and with the base frontend its success is lowest on stationary platforms (Section VI). We examined 27 terminal timeouts from a separate evaluation of this actor on slow or stationary platforms, selected because the horizontal pad error had fallen below 1 m. In 25 of them, the target was still at least partially visible at the first persistent loss of direct detection. Prompting SAM 3 with an oracle target box, which is unavailable onboard, recovered all 25 of these onset frames and 74 of 78 sampled frames in which the target was visible but undetected. Detection thus failed while the target could still be segmented. Later in these episodes, renderer-confirmed target absence followed. In the final 5 s, none of the 805 perception calls without a direct detection contained renderer-visible target pixels, and sustained renderer absence began a median of 2.08 s after the first persistent detection loss. This ordering is observational and does not show that the detection loss caused the later absence. It does show where target availability breaks down for the main actor, and it motivates the question of Section VI: whether the perception frontend can supply a usable target through these gaps without retraining the actor.
Temporal Memory Management
Persistent Tracking and Temporal Recovery
The base frontend bridges missed detections with one persistent tracking session per episode (Section III-B). After a long gap, that session continues from memory accumulated over earlier propagated frames. Temporal Recovery changes only how the session is anchored. With every fresh nonempty direct SAM 3 mask, it reinitialises the tracking session on the current frame, so that propagation always starts from the most recent direct detection. The output rule of (1), the observation format, and every learned weight stay the same, and the actor receives no tracker state, target position, or target velocity. The frontend is causal, and an episode reset clears the session, the CLIP template bank, and all counters. Figure 2 shows an example of a recovered target.

Figure 2. Example of temporal target recovery. A direct SAM 3 detection provides the earlier target observation (left). On a later frame without a direct detection (centre), the Temporal Recovery frontend propagates a target mask from the earlier visual evidence (right). The example is illustrative; closed-loop effects are reported in Section VI-B.
The Temporal Recovery wrapper also rejects a propagated mask whose native object-presence logit is not finite and positive. For finite scores this check repeats what DAM4SAM already does, because in the documented implementation DAM4SAM empties the propagated mask whenever the score is nonpositive. The effective difference from the base frontend is therefore reinitialisation at direct detections. This conclusion rests on a review of the source code (Appendix § 3.4); no closed-loop ablation separated reinitialisation from the wrapper check.
| Metric | Base | Temporal Recovery |
| Terminal usable-target coverage | 83.2% | 94.5% |
| Post-absence output | 17.9% | 30.1% |
Table 4. Target continuity and post-absence output of the two frontends, pooled over the 300 episodes per frontend of the first evaluation (Table V). Each frontend generates its own trajectories, so the values are descriptive. Terminal-coverage counts: 5,567/6,694 and 6,597/6,978 renderer-visible fresh frames; post-absence output counts: 430/2401 and 586/1945 fresh frames.
No predicted-IoU or CLIP threshold is applied to propagated masks.
Results
Temporal Recovery made the target more continuously available. Its terminal usable-target coverage exceeded that of the base frontend in all six conditions of the first evaluation, by 5.0 to 16.6 percentage points, and was 94.5% against 83.2% pooled over the evaluation (Table IV). The same reinitialised sessions also kept returning masks more often after renderer-confirmed target absence. After the last renderer-visible target frame of an episode, Temporal Recovery returned a nonempty mask on 30.1% of fresh frames, against 17.9% for the base frontend. Renderer absence can arise from a lateral exit from the field of view, clipping by the camera near plane, or other rendering effects, and the recorded traces do not establish their relative contributions. These rates are therefore the share of post-absence frames with a nonempty output. They are not counts of false detections, and a nonempty output on such a frame is not necessarily an error or a harmful control input.
Greater continuity did not translate into a reliable landing improvement. In the first evaluation (Table V), Temporal Recovery landed in 253 of 300 episodes and the base frontend in 237. Temporal Recovery was ahead in four of the six conditions and behind in S1 and S3. The largest difference was on stationary platforms (60% against 80%, ), and no per-condition difference reached the 0.05 level. The independent evaluation (Table VI) repeated the comparison on new seeds. Temporal Recovery landed in 88 of 100 nominal and 67 of 100 stationary episodes, against 83 and 63 for the base frontend. These differences are smaller than in the first
| Condition | Base | Temporal Recovery | SR (pp) | |
| P | 44/1/5 (88%) | 48/1/1 (96%) | +8 | 0.289 |
| S0 | 30/9/11 (60%) | 40/6/4 (80%) | +20 | 0.076 |
| S1 | 37/3/10 (74%) | 36/5/9 (72%) | −2 | 1.000 |
| S2 | 42/4/4 (84%) | 45/1/4 (90%) | +6 | 0.508 |
| S3 | 47/1/2 (94%) | 44/2/4 (88%) | −6 | 0.453 |
| S4 | 37/3/10 (74%) | 40/2/8 (80%) | +6 | 0.549 |
| Total | 237/21/42 (79.0%) | 253/17/30 (84.3%) | +5.3 | — |
Table 5. Landing outcomes of the frozen actor with the two frontends in the first evaluation, 50 matched episodes per condition. P and S1–S4 use separate seed blocks. Entries are success/crash/timeout (success rate). : exact two-sided McNemar test, without multiplicity correction.
| Cond. | Base | Temporal Rec. | SR (pp) | 95% CI (pp) | |
| P | 83/5/12 | 88/2/10 | +5.0 | [−4.73, +14.74] | 0.405 |
| S0 | 63/9/28 | 67/10/23 | +4.0 | [−9.02, +16.83] | 0.652 |
Table 6. Independent evaluation on 100 new matched episodes per condition. Entries are success/crash/timeout. The difference, its matched-pair 95% confidence interval (Newcombe square-and-add method with Wilson limits, pointwise and approximate), and the exact two-sided McNemar from a post hoc analysis that is conditional on the six unresolved pairs of Appendix B; the original protocol withheld paired inference.
evaluation, so the 48 of 50 nominal successes there are best read as a feature of that sample.
To express the remaining uncertainty, we analysed all 200 matched pairs of the independent evaluation post hoc. This analysis is conditional, because six pairs show slight platform-position differences before contact for reasons we could not identify (Appendix B). The differences of +5.0 and +4.0 percentage points have 95% confidence intervals of [−4.73, +14.74] and [−9.02, +16.83] points, so the data remain compatible with both a benefit and a harm. Excluding the six pairs leaves the direction unchanged, and the intervals still include zero. Temporal Recovery thus landed numerically more often, but the evaluations establish neither a reliable improvement nor equivalence or the absence of a benefit.
Discussion
The appearance evaluation shows that a frozen mask-conditioned landing policy can recover most of its performance on a substantially different pad when only the segmentation prompt is changed. This is adaptation through the prompt, not invariance to appearance: under the original prompt the banana-themed pad disrupted detection, and the policy landed in 12.0% of episodes. The mask interface moves the dependence on appearance from the controller to the perception model, where a text prompt can address it without new training data for the actor.
The results suggest that the availability of usable target masks was an important limiting factor in the tested conditions. In the appearance evaluation, landing success fell and recovered with the availability of direct detections. The substitution diagnostic, which used a different actor and protocol, points the same way. There, the nonempty masks were already geometrically accurate, and restoring missing masks for visible targets recovered landing, whereas the tested centroid and area corrections did not. Frame-level segmentation accuracy is therefore a weak proxy for the perception quality that this controller depends on, in line with task-aware evaluation of perception in other domains [17]. Recall on visible targets, especially partially visible targets at the image border, merits separate measurement when promptable segmentation is used as a control interface.
More availability did not simply mean more landings. Temporal Recovery raised terminal coverage, yet its effect on landing success remained uncertain. Several explanations are possible, and our experiments cannot distinguish between them. The base frontend already bridges gaps with a persistent tracker, so the additional coverage may matter in few episodes. Propagated masks carry their own source flag and may inform the actor less than direct detections. The sessions that bridge detection gaps also produce more nonempty output after renderer-confirmed target absence, which may offset part of the gain. Separating these effects would require configurations that were not evaluated here, such as a frontend without temporal fallback, one that holds the last direct mask, and a closed-loop contrast of reinitialisation alone. Temporal memory management thus trades missing observations against continued output after renderer-confirmed target absence, which may carry outdated target information. The appropriate degree of persistence depends on how the controller responds to each, which segmentation accuracy alone cannot determine. A variant that restricts propagation chains illustrates this trade-off (Appendix C).
Limitations
The appearance result concerns one substantially different pad, and the revised prompt was chosen by the authors after exploratory screening on development samples. Adaptation to arbitrary appearances and automatic prompt selection were not tested. The within-family textures are not new appearances, and the experiment does not separate the roles of colour, contrast, and semantics. The detector is a SAM 3 model fine-tuned for this project. Its original fine-tuning images were not available, so whether banana or related artwork appeared in its fine-tuning data could not be established (Appendix § 3.4), and its response to new appearances and prompts reflects this model and not the public release. Every frontend and appearance comparison uses one actor trained with one seed. The substitution diagnostic uses a separate policy, protocol, and observation interface, so its results do not transfer directly to the main actor, and ground-truth masks were not supplied to the main actor. The independent landing comparison rests on a post hoc analysis that is conditional on six unresolved pairs. All results come from simulation, with straight-line platform motion at constant speed and with ego velocity and attitude provided by the simulator. The causes of renderer-confirmed absence were not resolved, frontend latency was measured only on a server GPU, and the pipeline has not been flown.
Conclusion
We studied prompt-conditioned segmentation as the observation interface of a frozen visual RL policy for landing on a moving platform under a change of pad appearance. Revising the segmentation prompt recovered most of the landing performance lost under the tested appearance change without updating any network weight. An important limiting factor was whether the perception model delivered a target mask: landing followed the availability of direct detections, and in a controlled diagnostic, restoring missing masks mattered far more than correcting the tested geometric errors. Temporal tracking can supply masks through detection gaps, but it also increases nonempty output after renderer-confirmed target absence, and a landing improvement from greater continuity was not established. For segmentation-conditioned RL landing, target-mask availability, and the persistence with which a frontend supplies it, are the perception properties to measure and design for, within the scope of these simulation results.
Substitution Diagnostic Protocol
The substitution diagnostic of § 5 used a DrQ-v2 actor trained on renderer-derived ground-truth masks and the earlier protocol of Table VII. The confirmation used 200 paired episodes (seeds 2610000–2610199) after the 50-episode screen. The saved checkpoint of the ground-truth-mask actor (Table XII) records a global step counter of 299,890, which counts outer environment decisions and not physics steps, and a global episode counter of 3,097. The intended total training budget, the training seed, and the exact training-time source revision were not recovered. The raw-mask condition used the fine-tuned SAM 3 checkpoint of the main study with the prompt “landing pad”, thresholds of 0.5, an internal resolution of 1008 pixels, and no temporal fallback.
| Setting | Diagnostic | Main |
| Horizon / actor rate | 24 s / 6 Hz | 30 s / 12 Hz |
| Initial altitude | 0.8–6 m | 10 m |
| Horizontal offset | radial, | m/axis |
| Platform heading | ||
| Platform speed (m s) | 0 (30%), else | |
| Pad size | 0.675 m | 1 m |
| Horiz. velocity limit (m s) | 1.42 | 1.5 |
| Success criterion | altitude/contact | contact footprint |
| Low-dim. values per frame | 7 | 15 |
| SAM 3 checkpoint | same fine-tuned checkpoint | — |
| Temporal fallback | none | frontend-specific |
Table VII. Protocol of the substitution diagnostic, compared with the main protocol of Section III.
| Component | Configuration | SHA-256 |
| Actor | 1.25 M decisions, seed 1 | 926c61819980 |
| SAM 3 | fine-tuned, 1008 px | 120dc5228a36 |
| GT-mask actor | diagnostic only | 43bf505f11a5 |
| DAM4SAM | sam21pp_hiera_t.yaml | 7404e08d64fa |
| CLIP | ViT-B/32 (timm, OpenAI) | bd41409c7f2b |
| Temporal Recovery | runtime wrapper | ccea0ddabc1c |
Table XII. Frozen components (first 12 hexadecimal characters of each SHA-256 digest).
The diagnostic also used an earlier reward. Its low-dimensional input held seven ego-state values per frame; the main actor receives seven ego-state values and eight perception-metadata values per row of .
Independent Evaluation Statistics
Design. The independent evaluation used seeds 22300000–22300099 for P and 22310000–22310099 for S0, each run once with the base frontend and once with Temporal Recovery. All 400 episodes ran exactly once, with no replacement, retry, or technical failure. The weights of the frozen actor were verified to be identical in both arms.
Original pairing check and disposition. Matched episodes had to share their initial conditions, external platform motion commands, model and configuration identities, and simulation timing, whereas UAV actions, trajectories, outcomes, and durations may differ between frontends. The check compared the world-frame centre-of-mass position of the pad link of the platform vehicle between the two episodes. For each time-aligned post-motion physics step in their common recorded prefix, the difference is
the largest absolute coordinate difference, where TR denotes Temporal Recovery. Before the first qualifying UAV–vehicle contact in either episode, above 1 µm required review; from that contact onwards, differences were treated as possible physical responses. The check classified 159 pairs as within the threshold throughout, 35 as exceeding it only after contact, and 6 as exceeding it before contact (P: 78/17/5; S0: 81/18/1). Staying within the threshold does not mean exact numerical equality. Table VIII lists the six review cases. In all six, the recorded exogenous conditions, including resets, platform motion commands, model identities, and step and capture schedules, matched exactly, and the recorded state does not reveal what started the divergence. In five of them the threshold was first exceeded 2 to 15 physics steps before the first qualifying contact, so a near-contact physical response is plausible but unproven. In the sixth, seed 22300048, the first exceedance came about 9 s before the recorded qualifying contact, as isolated samples of at most 1.43 µm rather than a continuous excursion, with no established initiating cause. Because the original protocol required every pair to pass the check, the independent evaluation was originally reported with all 400 outcomes and without paired tests or confidence intervals. These six cases remain unresolved and are not shown to be harmless.
| Condition / seed | First step >1 µm | First contact step | Max. pre-contact difference (µm) |
| P / 22300015 | 5000 | 5015 | 3.80 |
| P / 22300036 | 5037 | 5039 | 1.49 |
| P / 22300047 | 4915 | 4917 | 1.55 |
| P / 22300048 | 4981 | 7196 | 1.43 |
| P / 22300098 | 4968 | 4970 | 1.31 |
| S0 / 22310096 | 4871 | 4874 | 1.35 |
Table VIII. Independent-evaluation pairs with unresolved pre-contact divergence. One physics step is 1/240 s.
Post hoc conditional reanalysis. A post hoc analysis of the existing records of all 200 matched pairs was later specified and frozen, after the outcomes and the six-case review were known but before any test or interval was computed. It is not a preregistered analysis and does not retroactively validate the original protocol. It rests on three assumptions: the recorded matching adequately defines a common scenario for each seed, unobserved state differences do not materially change the scenario for one arm only, and the seeds act as independent scenario replicates. Success differences were tested with exact two-sided McNemar tests on the discordant pairs. Pointwise 95% confidence intervals for the matched-pair difference in success probability were computed with Newcombe’s square-and-add method for paired data [27], using Wilson intervals for the two marginal rates. These intervals are approximate and are not inversions of the exact test. Table IX gives the results for all pairs and, as a sensitivity analysis, with the six review pairs excluded. This exclusion is itself post hoc and may introduce selection bias. With a Holm adjustment over the two conditions, both all-pair -values become approximately 0.810. In every analysis the difference favours Temporal Recovery, but every interval includes zero, so the data remain compatible with both benefits and harms. The reanalysis therefore establishes neither an improvement nor equivalence or the absence of a benefit.
| Analysis | Cond. | Pairs | Disc. | ΔSR (pp) | 95% CI (pp) | |
| All pairs | P | 100 | 9/14 | +5.00 | [−4.73, +14.74] | 0.405 |
| S0 | 100 | 20/24 | +4.00 | [−9.02, +16.83] | 0.652 | |
| Six excluded | P | 95 | 9/12 | +3.16 | [−6.70, +13.03] | 0.664 |
| S0 | 99 | 19/24 | +5.05 | [−7.97, +17.83] | 0.542 |
Table IX. Post hoc conditional reanalysis of the independent evaluation: Temporal Recovery minus base frontend. Disc.: discordant pairs, successes with the base frontend only / with Temporal Recovery only.
Other evaluations. The first evaluation and the appearance evaluation also use matched episodes, and their reset states were verified, but the pre-contact platform-position check above was not applied to them retrospectively. Two earlier runs of the base versus Temporal Recovery comparison, a 40-episode pilot and a run stopped before its planned sample was complete, are not analysed in this paper.
Memory-Restricted Tracking
Memory-based video segmentation conditions later predictions on stored temporal information, so an erroneous propagated mask can affect subsequent frames [16]. In Temporal Recovery, a session reanchored at a direct detection keeps propagating through consecutive misses, so its later predictions depend on earlier propagated masks. The memory-restricted variant keeps as its anchor the RGB image and mask of the most recent fresh direct detection, which only another direct detection refreshes. On each fresh frame without a direct detection, a new tracking session is initialised from this anchor and propagated once to the current frame, so propagated predictions never become anchors for later frames. The restriction is implemented by resetting sessions in the wrapper; it does not reproduce the method of [16], no model weight changes, and the observation format and cadence are those of Temporal Recovery.
Screening. Before any closed-loop evaluation, the variant was screened against a previously set internal engineering criterion of at most six false target reports on a fixed set of 54 identical frames, half of the 12 reports of Temporal Recovery on the same frames. The criterion was not a statistical threshold or a physically derived limit. The variant produced seven reports and did not meet it. We report its closed-loop results below descriptively; the screening result alone does not determine their interpretation.
Observations. In the first evaluation (Table X), the variant landed in 251 of 300 episodes, against 253 for Temporal Recovery and 237 for the base frontend (Table V), with a different profile across speeds. Its terminal usable-target coverage was 90.8% (6,321/6,958 renderer-visible fresh frames), between those of the base frontend (83.2%) and Temporal Recovery (94.5%), and its post-absence output rate was 12.8% (238/1855 fresh frames), the lowest of the three. When conditions A, B, and C of the appearance evaluation were run with this variant on the same matched episodes, success was 78.0%, 8.8%, and 66.8%, the same pattern as with Temporal Recovery; matched comparisons with Temporal Recovery gave , 0.280, and 0.734, which do not show that the two frontends are equivalent. The variant illustrates the persistence trade-off of Section VII: restricting propagation lowered post-absence output at some cost in coverage. It is not proposed as a method, and these results do not show a landing improvement.
| Condition | Memory-restricted |
| P | 41/2/7 (82%) |
| S0 | 44/1/5 (88%) |
| S1 | 43/2/5 (86%) |
| S2 | 38/5/7 (76%) |
| S3 | 44/1/5 (88%) |
| S4 | 41/2/7 (82%) |
| Total | 251/13/36 (83.7%) |
Table X. Landing outcomes of the memory-restricted variant in the first evaluation, 50 matched episodes per condition, matched with Table V. Entries are success/crash/timeout (success rate).
Appendix D: Implementation and Reproducibility
Policy and training. The DrQ-v2 encoder consists of four unpadded convolutions with 32 channels and strides of 2, 1, 1, and 1, each followed by a ReLU, and yields 39,200 visual features, which are concatenated with the 45 entries of the flattened history. Between perception updates the mask history is held; on the intermediate actor decision, the ego-state entries and the detection age of the newest row are refreshed. The critic receives the platform position and velocity minus those of the UAV in the world frame, normalised by m and m s. The encoder is updated only by critic gradients, the actor receives its output detached, and the critic passes the features and the six privileged scalars through its own trunk of the same form and two Q heads with two 512-unit hidden layers each. The reward is
with the potential
of the horizontal distance , the horizontal relative speed , and the positive relative altitude . For terminal transitions is set to zero. The decision interval is normally s, but an early contact within the first PID substep can give s. The discount per actor transition keeps the physical-time discount of an earlier 6 Hz version of the actor, which used 0.99 per decision. Training ran for 500,000 curriculum decisions, starting at 2 m with offsets in m and platform speeds in m s for the first 100,000 decisions, then at 5 m with m and m s until 250,000, and then under the nominal distribution. A further 500,000 decisions used the nominal distribution, and a final 250,000 sampled half of the episodes from the nominal speed distribution and a quarter each from and m s. We did not establish what each phase contributes. Table XI lists the settings.
| Setting | Value |
| Training (DrQ-v2) | |
| Total actor decisions | |
| Mask stack | |
| Low-dimensional history | 3 |
| Optimiser / learning rate | Adam / |
| Batch size / replay capacity | 256 / 160,000 |
| Discount | 0.99498743710662 |
| Target-update | 0.05 |
| -step return horizon / update interval | 6 / 2 decisions |
| Random / update warm-up | 2,000 / 4,000 decisions |
| Training seed | 1 |
| Exploration | 1.0 0.1 over |
| Update noise clip | 0.3 |
| Random-shift padding | 4 pixels |
| Brightness multiplier | |
| Reward | |
| Success / crash / timeout | +140 / −50 / −45 |
| Time cost | −0.05 |
| Horizontal-distance weight / scale | 4 / 3 m |
| Relative-horizontal-speed weight / scale | 0.5 / 1 m s |
| Altitude weight / scale | 1 / 10 m |
| Terminal next-state potential | 0 |
Table XI. Training and reward settings of the main actor.
SAM 3 fine-tuning. The SAM 3 detector was fine-tuned for this project in the run landing_pad_generalized_v2_accum16. The archived fine-tuning pipeline specifies a single COCO category, “landing pad”, which is also the text query, and documents a randomised landing-scene dataset with 400 training and 80 validation image entries, of which 40 and 8 are intended as negative examples without a target. These are documented dataset specifications, not verified counts of unique images used in the final run: the original training images, masks, and manifests were not available, and how the masks were produced could not be established. The recorded configuration specifies 10 epochs, a model input resolution of 1008 pixels, the AdamW optimiser, and frozen vision and language backbones. Whether banana or related artwork appeared in the fine-tuning data could not be established from the records. The first author confirms that no banana-themed pad images or banana-related prompts were deliberately included; this is an author statement, not a verification against retained data. No statement is made about the large-scale pretraining of the public SAM 3 model, whose exact upstream release revision is not available. After fine-tuning, the checkpoint was frozen during reinforcement learning and all evaluations.
Prompt selection. Five candidate prompts were compared on the same five fixed views of the banana-themed pad by scoring the resulting SAM 3 masks against the rendered pad, without running any landing episode: “landing pad”, “square landing pad”, and “the full square landing target” returned no mask on any view, “landing pad with banana print” gave two whole-pad masks, one partial mask, and two empty outputs, and “banana-printed landing pad” gave four whole-pad masks and one empty output. No prompt produced an acceptable mask on all five views, as the screen required, and retests with the revised banana texture again gave four acceptable views and one empty output. The prompt and texture were then fixed, before any landing episode was run with them. With the prompt fixed, the same conditions were first evaluated with the memory-restricted frontend, on a block of 375 development episodes and then on a new set of 250 seeds. The Temporal Recovery evaluation of Section IV came afterwards and reused this 250-seed set to match the two frontends, so its episodes had already been evaluated with the memory-restricted frontend. These seeds overlap neither the fixed views of the prompt comparison nor the development episodes. The available records do not indicate that the Temporal Recovery landing outcomes were used to choose the prompt, but they do not exclude undocumented prompt attempts.
Perception details. The CLIP score is the mean cosine similarity between the RGB bounding-box crop of the current output mask and the two most similar templates in a causal bank of crops from direct detections, or the similarity to the only template when the bank holds one; it uses no text embedding. The rendered target mask used by the perception metrics comes from the body identifier of the platform vehicle, including its links. In the documented DAMASAM implementation, a nonpositive object-presence score replaces the predicted mask logits before thresholding. In the independent evaluation, 2,881 perception frames across 178 of the 200 Temporal Recovery episodes had a nonpositive presence score, and every one produced an empty output. Latency was timed synchronously on four recorded RGB sequences after 20 warm-up calls per component; 2 of 585 base-frontend updates and 1 of 585 Temporal Recovery updates exceeded the 166.7 ms interval, and peak allocated GPU memory was 4.60 and 4.73 GiB.
Frozen components and runtime. Table XII lists the frozen components. The full SHA-256 digests of the SAM 3 and ground-truth-mask actor checkpoints are 120dc5228a36cd1b1cb841a7607609c4b4f0cad30c1e3ad76c88cfd94cf30891 and 43bf505f11a5c96ced7b5fe384e6ec5e e128185e9187ff65de2bede8d6e4ce70. The recorded runtime environment used the optional compiled post-processing fallback of SAM 2, and the independent evaluation logged a texture-loading warning for both arms of seed 22310099 (S0).
Acknowledgment
The first author used AI assistants during this work. Chat-GPT and Codex (OpenAI) mainly assisted with writing code, and Claude (Anthropic) helped polish its language and grammar. The experiments, results, and scientific decisions are the authors’ own, and the authors checked all text, numbers, and references and take full responsibility for the content of this paper.
References
- [1]P. Goldschmid and A. Ahmad, “Reinforcement learning based autonomous multi-rotor landing on moving platforms,” Autonomous Robots, vol. 48, Art. no. 13, 2024, doi: 10.1007/s10514-024-10162-8.
- [2]D. Falanga, A. Zanetti, A. Simovic, J. Delmerico, and D. Scaramuzza, “Vision-based autonomous quadrotor landing on a moving platform,” in 2017 IEEE International Symposium on Safety, Security and Rescue Robotics (SSRR), Shanghai, China, 2017, pp. 200–207, doi: 10.1109/SSRR.2017.8088164.DOI
- [3]A. Paris, B. T. Lopez, and J. P. How, “Dynamic Landing of an Autonomous Quadrotor on a Moving Platform in Turbulent Wind Conditions,” in 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 2020, pp. 9577–9583, doi: 10.1109/ICRA40945.2020.9197081.DOI
- [4]P. Ladosz, M. Mammadov, H. Shin, W. Shin, and H. Oh, “Autonomous Landing on a Moving Platform Using Vision-Based Deep Reinforcement Learning,” IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4575–4582, May 2024, doi: 10.1109/LRA.2024.3379837.DOI
- [5]G. Cho, J. Choi, G. Bae, and H. Oh, “Autonomous ship deck landing of a quadrotor UAV using feed-forward image-based visual servoing,” Aerospace Science and Technology, vol. 130, Art. no. 107869, 2022, doi: 10.1016/j.ast.2022.107869.
- [6]B. Herissé, T. Hamel, R. Mahony, and F.-X. Russotto, “Landing a VTOL Unmanned Aerial Vehicle on a Moving Platform Using Optical Flow,” IEEE Transactions on Robotics, vol. 28, no. 1, pp. 77–89, Feb. 2012, doi: 10.1109/TRO.2011.2163435.
- [7]P. Serra, R. Cunha, T. Hamel, D. Cabecinhas, and C. Silvestre, “Landing of a Quadrotor on a Moving Target Using Dynamic Image-Based Visual Servo Control,” IEEE Transactions on Robotics, vol. 32, no. 6, pp. 1524–1535, Dec. 2016, doi: 10.1109/TRO.2016.2604495.
- [8]A. Rodriguez-Ramos, C. Sampedro, H. Bavle, P. de la Puente, and P. Campoy, “A Deep Reinforcement Learning Strategy for UAV Autonomous Landing on a Moving Platform,” Journal of Intelligent & Robotic Systems, vol. 93, no. 1–2, pp. 351–366, 2019, doi: 10.1007/s10846-018-0891-8.
- [9]R. Polvara, M. Patacchia, M. Hanheide, and G. Neumann, “Sim-to-Real Quadrotor Landing via Sequential Deep Q-Networks and Domain Randomization,” Robotics, vol. 9, no. 1, Art. no. 8, 2020, doi: 10.3390/robotics9010008.
- [10]D. Yarats, R. Fergus, A. Lazaric, and L. Pinto, “Mastering Visual Continuous Control: Improved Data-Augmented Reinforcement Learning,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2022.
- [11]A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, P. Dollár, and R. Girshick, “Segment Anything,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), Paris, France, 2023, pp. 3992–4003, doi: 10.1109/ICCV51070.2023.00371.
- [12]N. Carion et al., “SAM 3: Segment Anything with Concepts,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2026.arxiv.org/abs/2511.16719
- [13]W. Shin, M. Kim, T. Park, G. Bae, S. Kim, and H. Oh, “Vision-Based Autonomous Drone Landing on Moving Platforms With Uncertain Motion via Deep Reinforcement Learning,” IEEE Robotics and Automation Letters, vol. 11, no. 5, pp. 5542–5549, May 2026, doi: 10.1109/LRA.2026.3674011.
- [14]G. Aikins, S. Jagtap, and K.-D. Nguyen, “A Robust Strategy for UAV Autonomous Landing on a Moving Platform under Partial Observability,” Drones, vol. 8, no. 6, Art. no. 232, 2024, doi: 10.3390/drones8060232.
- [15]N. Ravi et al., “SAM 2: Segment Anything in Images and Videos,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2025.
- [16]S. Ding, R. Qian, X. Dong, P. Zhang, Y. Zang, Y. Cao, Y. Guo, D. Lin, and J. Wang, “SAM2Long: Enhancing SAM 2 for long video segmentation with a training-free memory tree,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 13614–13624.
- [17]J. Philion, A. Kar, and S. Fidler, “Learning to evaluate perception models using planner-centric metrics,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 14055–14064.
- [18]L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asymmetric Actor Critic for Image-Based Robot Learning,” in Proc. Robotics: Science and Systems (RSS), 2018.
- [19]H. K. Cheng and A. G. Schwing, “XMem: Long-Term Video Object Segmentation with an Atkinson–Shiffrin Memory Model,” in Computer Vision – ECCV 2022, Lecture Notes in Computer Science, vol. 13688, Springer, Cham, 2022, pp. 640–658, doi: 10.1007/978-3-031-19815-1_37.
- [20]J. Panerati, H. Zheng, S. Zhou, J. Xu, A. Prorok, and A. P. Schoellig, “Learning to Fly—A Gym Environment with PyBullet Physics for Reinforcement Learning of Multi-agent Quadcopter Control,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Prague, Czech Republic, 2021, pp. 7512–7519, doi: 10.1109/IROS51168.2021.9635857.
- [21]A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proc. 38th Int. Conf. Mach. Learn. (ICML), PMLR vol. 139, 2021, pp. 8748–8763.
- [22]J. Videnović, A. Lukežič, and M. Kristan, “A distractor-aware memory for visual object tracking with SAM2,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 24255–24264.
- [23]A. Y. Ng, D. Harada, and S. Russell, “Policy Invariance Under Reward Transformations: Theory and Application to Reward Shaping,” in Proc. 16th Int. Conf. Mach. Learn. (ICML), 1999, pp. 278–287.
- [24]M. Müller, A. Dosovitskiy, B. Ghanem, and V. Koltun, “Driving Policy Transfer via Modularity and Abstraction,” in Proc. 2nd Conf. Robot Learn. (CoRL), PMLR vol. 87, 2018, pp. 1–15.
- [25]B. Zhou, P. Krähenbühl, and V. Koltun, “Does Computer Vision Matter for Action?” Science Robotics, vol. 4, no. 30, Art. no. eaaw6661, 2019, doi: 10.1126/scirobotics.aaw6661.DOI
- [26]Z. Wang, Y. Ze, Y. Sun, Z. Yuan, and H. Xu, “Generalizable Visual Reinforcement Learning with Segment Anything Model,” arXiv:2312.17116, 2023.
- [27]R. G. Newcombe, “Improved confidence intervals for the difference between binomial proportions based on paired data,” Stat. Med., vol. 17, no. 22, pp. 2635–2650, 1998, doi: 10.1002/(SICI)1097-0258(19981130)17:22<2635::AID-SIM954>3.0.CO;2-C.