betaFeedback

WorksComputer Science

Control Barrier Functions for Safety-Aware Latent-Space Motion Planning

Abstract

Latent-space path planning (LSPP) compresses the configuration space of a robotic manipulator into a compact, differentiable manifold learned by a variational autoencoder, enabling fast gradient-based planning without explicit inverse kinematics. Existing LSPP frameworks enforce obstacle avoidance with a binary collision classifier, whose saturated gradients near obstacle boundaries and lack of temporal structure cause the optimizer to stall and provide no forward-invariance guarantee. We replace the classifier with a neural control barrier function (CBF) learned directly in the latent space. The barrier is trained with a three-term objective that combines safe and unsafe sign constraints with a discrete-time decrease condition evaluated on transitions of an obstacle-blind nominal planner, and it is enforced at planning time by a closed-form, first-order correction that projects each goal-directed latent step back onto the CBF constraint. On a 7-DoF Franka Panda manipulator, the CBF planner achieves a 75.6% success rate on a fixed-obstacle task, 12.0 percentage points above the best of 35 tuned classifier configurations, and 83.8% on a multi-position task. A sweep of the decay rate αα characterizes the safety–performance trade-off, with the best operating points in α∈[0.10,0.20]α\in [0.10, 0.20]. In a fully general random-obstacle setting, success drops to 33.0%, which we trace to covariate shift between the obstacle-blind training transitions and the states visited by the corrected planner.

Introduction

Motion planning for robotic manipulators in high-dimensional configuration spaces remains computationally demanding. A 7-degree-of-freedom (DoF) manipulator such as the Franka Emika Panda [1] must reach a goal while satisfying joint limits, avoiding workspace obstacles, and remaining kinematically feasible (Figure 1). Safe and reliable operation is essential in safety-critical domains such as healthcare [2] and bipedal locomotion [3]. Sampling-based planners such as Rapidly-exploring Random Trees (RRT) and Probabilistic Roadmaps (PRM) are probabilistically complete [4], but their runtime scales poorly with dimension. Optimization-based planners such as CHOMP [5] and TrajOpt [6] scale better, but they are sensitive to initialization, prone to local minima, and rely on expensive distance-field queries or exact collision checks at every iteration. Model predictive control (MPC) [7] offers a principled receding-horizon alternative but faces the same dimensionality.

Top: A naive trajectory executed without collision avoidance intersects the obstacle. Bottom: A safe trajectory guides the manipulator over the obstacle to the target.

Figure 1. Top: A naive trajectory executed without collision avoidance intersects the obstacle. Bottom: A safe trajectory guides the manipulator over the obstacle to the target.

Learned latent representations offer an alternative by compressing the planning problem into a lower-dimensional, smooth manifold. A variational autoencoder (VAE) [8, 9] trained on valid robot configurations learns a compact latent space that implicitly captures kinematic feasibility and the

multimodality of inverse kinematics. Gradient-based optimization in this space, termed activation maximization (AM), can then pursue goal reaching and obstacle avoidance jointly. Hung et al. [10] instantiate this idea for manipulation, building on the broader paradigm of planning in learned generative manifolds [11]: a VAE encodes joint configurations and end-effector positions, and a learned binary collision classifier steers the latent state away from predicted collisions.

Relying on a binary classifier for obstacle avoidance has two structural weaknesses. First, its sigmoid output yields a poor gradient landscape: far from an obstacle the gradient is nearly zero and provides no anticipatory guidance, while near or inside the obstacle it spikes. This “cliff-edge” profile traps the trajectory optimizer in local minima. Second, a classifier evaluates static configurations and has no temporal structure, so nothing prevents a state judged safe at step kk from moving into collision at step k+1k+1.

We address both weaknesses by learning a control barrier function (CBF) [12, 13] directly in the VAE latent space. A CBF B(z)B(z) is a continuous function whose nonnegative superlevel set is the safe set; its decrease condition bounds how fast BB may decay from one step to the next, which renders the safe set forward invariant. Rather than adding a soft collision penalty to the planning objective, we use the learned barrier as a closed-form safety filter: each goal-directed latent step is computed without regard to obstacles and, if it violates the CBF condition, is projected along the barrier gradient back onto the constraint. Neural CBFs have emerged as a data-driven way to synthesize such certificates [14], but their integration into learned generative planning spaces remains underexplored. Our contributions are:

  • A reproduction of the LSPP baseline [10] for a 7-DoF Franka Panda, with a two-stage evaluation that validates every plan by exact geometric collision checking.

  • A neural CBF defined in the learned latent space and trained with a three-term objective (safe sign, unsafe sign, and discrete decrease condition) on a state-label dataset and on transitions of an obstacle-blind nominal planner.

  • A closed-form, first-order safety correction that filters each latent planning step through the CBF constraint, decoupling safety from goal pursuit.

  • An evaluation of three barrier variants of increasing generality, B(z)B(z), B(z,x,y)B(z,x,y), and B(z,o)B(z,o), with a systematic characterization of the α\alpha safety–performance trade-off. The CBF outperforms the best-tuned classifier by 12.0 percentage points on a fixed-obstacle task and reaches 83.8% success on a multi-position task; in the general setting we identify covariate shift as the main failure mode.

Section II reviews related work and Section III presents the background and baseline framework. Section IV introduces the latent-space CBF, Sections V and VI describe the experimental setup and results, and Sections VII and VIII discuss limitations, future work, and conclusions.

Related Work

Motion Planning and Learned Representations

Sampling-based algorithms avoid constructing the configuration space explicitly, but in high dimensions they yield highly variable planning times and jerky paths that require post-processing. Trajectory optimization minimizes a cost functional combining smoothness and obstacle avoidance [5], [6], but complex geometry is hard to represent analytically and non-convex environments trap the optimizer. Both paradigms rely on forward kinematics and geometric intersection tests inside the planning loop. Learning from demonstrations offers a data-driven alternative, including behavioral cloning [15], inverse reinforcement learning [16], generative adversarial imitation [17], and goal-conditioned imitation that generalizes across targets [18]. Generative models enable planning in learned manifolds [10], [11], and diffusion models have been used directly as trajectory planners [19]. We keep the latent planner of [10] and replace its safety mechanism.

Control Barrier Functions and Neural Certificates

CBFs were introduced in the context of adaptive cruise control [12] and formalized in [13], [20]. They have since been applied to UAV flight [21], [22], bipedal walking [3], and quadrotor control [23]. When analytical barriers are unavailable, they can be learned from data: Robey et al. [24] learn barrier certificates from expert demonstrations, and Srinivasan et al. [25] enforce control-theoretic inequalities as empirical losses over sampled states. Dawson et al. [14], [23] learn neural Lyapunov-barrier functions with formal guarantees. Exact verification of ReLU-based neural CBFs [26], automated certificate synthesis with FOSSIL [27], training under input constraints [28], and conformal prediction for verified neural CBFs [29] address the reliability of learned certificates, and physics-informed objectives can improve them further [30]. Hamilton–Jacobi (HJ) reachability [31] characterizes backward reachable sets exactly; Fisac et al. [32] connect HJ safety analysis with reinforcement learning, and robust control barrier-value functions [33] combine the CBF and HJ perspectives.

Safe Learning from Data

Surveys of safe reinforcement learning (RL) [34], [35] group approaches into constrained optimization [36], [37], shielding [38], and barrier-based methods, building on standard RL foundations [39]. Offline RL learns from fixed datasets [40], where extrapolation error [41] and bootstrapping error [42] motivate conservative [43] and implicit [44] value learning. In offline safe RL, feasibility-guided diffusion [45], reachability-based flow policies [46], and epigraph-guided flow matching [47] show that expressive generative policies can balance safety and task performance, while value-guided offline CBFs learn safety filters entirely from offline data [48]. Safety has also been built into imitation learning through end-to-end imitation with CBF guarantees [49], in-distribution barrier functions [50], robust imitation through stochastic encoding [51], safety-guided imitation [52], soft barrier certificates for unknown stochastic environments [53], and semi-supervised safe visuomotor policies [54].

Background and Baseline Framework

Problem Setting

Given a start configuration qstartq_{\mathrm{start}} and a target end-effector position etargete_{\mathrm{target}}, the planner must find a continuous trajectory τ(t)∈Cfree\tau(t) \in\mathcal{C}_{\mathrm{free}} that ends at a configuration etargete_{\mathrm{target}}, where Cfree\mathcal{C}_{\mathrm{free}} is the collision-free subset of the configuration space. Following [10], the state of the Franka Panda is x=(q,e)∈R10x=(q,e) \in\mathbb{R}^{10}, where q∈R7q \in\mathbb{R}^{7} are the joint angles and e∈R3e \in\mathbb{R}^{3} is the end-effector position given by forward kinematics. Including ee in the state lets the planner target positions directly without solving inverse kinematics; the relationship between qq and ee is learned implicitly. All ten dimensions are standardized with training-set statistics, and the collision classifier reuses the statistics of the free-space dataset so that it operates in the same normalized space as the VAE. Obstacles are cylinders described by o=[x,y,h,r]o=[x,y,h,r], their planar position, height, and radius.

Latent Representation

The VAE maps x∈R10x \in\mathbb{R}^{10} to a latent code z∈R7z \in\mathbb{R}^{7} (Figure 2). The encoder has four hidden layers of 2048 units with exponential linear unit (ELU) activations and outputs μ∈R7\mu\in\mathbb{R}^{7} and σ2∈R7\sigma^{2} \in\mathbb{R}^{7}, with σ2=softplus⁡(⋅)+10−5\sigma^{2}=\operatorname{softplus}(\cdot)+10^{-5} for numerical stability; samples are drawn as z=μ+ϵ⊙σz = \mu+ \epsilon\odot\sigma with ϵ∼N(0,I)\epsilon\sim\mathcal{N}(0, I). The decoder mirrors the encoder and outputs x^=(q^,e^)\hat{x} = (\hat{q}, \hat{e}).

Baseline LSPP architecture

Figure 2. Baseline LSPP architecture. Top: The VAE compresses the robot state (q∈R7,e∈R3)(q \in\mathbb{R}^{7}, e \in\mathbb{R}^{3}) into a 7-D latent space regularized by a prior. Bottom: The binary collision classifier predicts collision from the latent code concatenated with the obstacle descriptor o∈R4o \in\mathbb{R}^{4}.

The standard evidence lower bound (ELBO) [8] balances reconstruction against a Kullback–Leibler (KL) penalty toward the prior p(z)=N(0,I)p(z) = \mathcal{N}(0, I), and β\beta-VAE [9] modulates this weight to obtain more disentangled representations. Neither guarantees the strict reconstruction fidelity that robot kinematics requires, so we use the Generalized ELBO with Constrained Optimization (GECO) [55], which treats reconstruction quality as a constraint and minimizes

LGECO=DKL[qϕ(z∣x)∥p(z)]+λEz[max⁡(0,MSE⁡(x,x^)−ggoal)],(1)\mathcal{L}_{\mathrm{GECO}} = D_{\mathrm{KL}}\left[q_{\phi}(z \mid x) \mathbin{\|} p(z)\right] + \lambda\mathbb{E}_{z}\left[\max\left(0, \operatorname{MSE}(x,\hat{x}) - g_{\mathrm{goal}}\right)\right], \tag*{(1)}

where the Lagrange multiplier λ\lambda is updated by an exponential moving average: it grows while the mean squared error exceeds the threshold ggoalg_{\mathrm{goal}} and shrinks once the constraint is met, letting the network prioritize compression. We train for 16,000 epochs with Adam (learning rate 10−410^{-4}, batch size 4096) and ggoal=0.0008g_{\mathrm{goal}} = 0.0008. The checkpoint is selected by the area under the curve of an AM planning metric rather than by validation loss, so that the chosen model supports stable gradient-based planning.

Collision Classifier

In the baseline, a classifier head is appended to the frozen VAE encoder. It takes the 11-D input (z,o)(z,o), has four hidden layers of 2048 units, and outputs a logit ℓ(z,o)\ell(z,o) whose sigmoid gives P(collision∣z,o)P(\mathrm{collision} \mid z,o). It is trained with binary cross-entropy (BCE) for 16,000 epochs.

Activation Maximization Planning

AM performs gradient descent on zz to minimize

LAM(z)=∥e^(z)−etarget∥2+λprior2∥z∥2−λobs∑ilog⁡(1−σ(ℓ(z,oi)/T)).(2)\mathcal{L}_{\mathrm{AM}}(z) = \lVert\hat{e}(z) - e_{\mathrm{target}}\rVert^{2} + \frac{\lambda_{\mathrm{prior}}}{2}\lVert z\rVert^{2} - \lambda_{\mathrm{obs}} \sum_i \log\left(1-\sigma\left(\ell(z,o_i)/T\right)\right). \tag*{(2)}

The first term drives the end effector toward the target, the second keeps zz near the VAE prior to avoid out-of-distribution, kinematically invalid states, and the third minimizes collision probability, with temperature T=3.0T = 3.0 softening the classifier gradients. The optimizer is initialized with a sample of the encoder distribution at xstartx_{\mathrm{start}} rather than its mean, which provides exploration noise in the non-convex loss landscape. Planning runs Adam with learning rate 0.15 for at most 300 steps; each iterate is decoded into a waypoint (q^k,e^k)(\hat{q}_k,\hat{e}_k), and planning stops once ∥e^k−etarget∥<10 mm\lVert\hat{e}_k-e_{\mathrm{target}}\rVert< 10\ \mathrm{mm}.

Control Barrier Functions

For a system x˙=f(x,u)\dot{x}=f(x,u), the safe set C={x:B(x)≥0}\mathcal{C}=\{x:B(x)\geq0\} is the superlevel set of a continuously differentiable function BB [13, 20]. C\mathcal{C} is forward invariant if there exists an extended class-K\mathcal{K} function α\alpha such that

B˙(x)+α(B(x))≥0.(3)\dot{B}(x)+\alpha(B(x))\geq0. \tag*{(3)}

For discrete iterates with step Δt\Delta t and a scalar decay rate α\alpha satisfying 0≤αΔt≤10\leq\alpha\Delta t\leq1, the corresponding condition is

B(zk+1)≥(1−αΔt)B(zk),(4)B(z_{k+1})\geq(1-\alpha\Delta t)B(z_k), \tag*{(4)}

which bounds the fraction of the safety margin that may be lost in a single step.

Latent-Space Control Barrier Functions

We replace the classifier term of Eq. (2) with a neural barrier B(z,o)B(z,o) trained in the latent space and enforced through a predictor–corrector step. The design draws on neural CBF synthesis [14, 23, 28], safety-guided imitation [49, 52], and in-distribution barrier functions [50].

Why a Barrier Instead of a Classifier

Gradient quality. The BCE objective saturates for confident predictions. A barrier trained as an unbounded, signed function instead provides dense, non-vanishing gradients ∇zB(z,o)\nabla_z B(z,o) that reflect the local boundary geometry, an advantage also observed for CBF-based obstacle avoidance on UAVs [22].

Temporal structure. The discrete condition (4) couples zkz_k and zk+1z_{k+1} and rules out latent steps that remove most of the safety margin in a single iteration.

Modularity. In the baseline, goal seeking and safety compete within one scalar loss. With a CBF, the nominal goal-directed step is computed freely and safety is applied afterwards as a projection, which avoids gradient interference. This mirrors the separation of policy and safety filter in safe visuomotor control [54].

Training Data

The barrier must capture both spatial geometry (where obstacles are) and the planner’s dynamics (how latent iterates move). Analytical obstacle boundaries are intractable in the non-linear 7-D latent space, so we learn the barrier from two datasets (Algorithm 1):

  • State-label dataset Dstate\mathcal{D}_{\mathrm{state}}: collision-labeled configurations encoded deterministically through the VAE mean into tuples (z,o,label)(z,o,\mathrm{label}), augmented with samples near obstacle surfaces to sharpen the decision boundary.

Algorithm 1: Dual-dataset generation for CBF training
Input: Collision dataset $\mathcal{D}_{\mathrm{base}}$, nominal planner, geometric simulator
Output: $\mathcal{D}_{\mathrm{state}}$, $\mathcal{D}_{\mathrm{trans}}$
$\mathcal{D}_{\mathrm{state}} \leftarrow \varnothing$, $\mathcal{D}_{\mathrm{trans}} \leftarrow \varnothing$;
foreach $(q,o,\mathrm{label}) \in \mathcal{D}_{\mathrm{base}}$ do
  $\mu,\sigma \leftarrow \operatorname{Encoder}(q)$;
  $\mathcal{D}_{\mathrm{state}} \leftarrow \mathcal{D}_{\mathrm{state}} \cup \{(\mu,o,\mathrm{label})\}$;
  for $i=1$ to $N_{\mathrm{scenarios}}$ do
    Sample $z_0 \sim \operatorname{Encoder}(q_{\mathrm{start}})$;
    for $k=0$ to $K_{\max}$ do
      // obstacle-blind nominal step
      $z_{\mathrm{nom}} \leftarrow z_k-\operatorname{Adam}(\nabla(\mathcal{L}_{\mathrm{goal}}+\mathcal{L}_{\mathrm{prior}}))$;
      $l_k \leftarrow \operatorname{Simulator.safe}(z_k)$;
      $l_{\mathrm{nom}} \leftarrow \operatorname{Simulator.safe}(z_{\mathrm{nom}})$;
      $\mathcal{D}_{\mathrm{trans}} \leftarrow \mathcal{D}_{\mathrm{trans}} \cup \{(z_k,z_{\mathrm{nom}},o,l_k,l_{\mathrm{nom}})\}$;
      $z_{k+1} \leftarrow z_{\mathrm{nom}}$;
return $\mathcal{D}_{\mathrm{state}},\mathcal{D}_{\mathrm{trans}}$
  • **Transition dataset Dtrans∗∗:pairs\mathcal{D}_{\mathrm{trans}}**: pairs(z_k,z_{\mathrm{nom}})$ produced by a nominal planner that optimizes only the goal and prior terms and ignores obstacles, with each state labeled by a geometric simulator. Because the barrier will correct exactly this obstacle-blind optimizer at inference time, it must learn the decrease condition on that optimizer’s transitions.

This follows Robey et al. [24], who learn barrier certificates from sampled trajectories when analytical models are unavailable, and Srinivasan et al. [25], who enforce control-theoretic inequalities as empirical losses; here, the nominal planner’s transitions play the role of expert trajectories.

Barrier Network and Training Objective

The barrier network takes the 11-D input (z,o)(z,o), has four hidden layers of 2048 ELU units, and outputs a scalar B∈RB \in\mathbb{R} without a final sigmoid; B≥0B \ge0 denotes the safe set. It is trained to minimize

LCBF=λsEsafe[ReLU⁡(γ−B(z,o))]+λuEunsafe[ReLU⁡(B(z,o)+γ)]+λdEtrans[ReLU⁡(Btarget−B(znom,o))],(5)\mathcal{L}_{\mathrm{CBF}} = \lambda_s \mathbb{E}_{\mathrm{safe}}[\operatorname{ReLU}(\gamma-B(z,o))] + \lambda_u \mathbb{E}_{\mathrm{unsafe}}[\operatorname{ReLU}(B(z,o)+\gamma)] + \lambda_d \mathbb{E}_{\mathrm{trans}}[\operatorname{ReLU}(B_{\mathrm{target}}-B(z_{\mathrm{nom}},o))], \tag*{(5)}

where Btarget=(1−αΔt)B(zk,o)B_{\mathrm{target}}=(1-\alpha\Delta t)B(z_k,o). The first two terms are evaluated on Dstate\mathcal{D}_{\mathrm{state}} and the third on Dtrans\mathcal{D}_{\mathrm{trans}}. The margin γ=1\gamma=1 requires B≥1B \ge1 on safe states and B≤−1B \le-1 on unsafe states; this two-unit gap buffers against VAE decoding variance. To isolate the difficulty of obstacle generalization, we train three variants:

  • FixedBarrierNet B(z)B(z): 7-D input, one fixed obstacle.

  • MultiPosBarrierNet B(z,x,y)B(z,x,y): 9-D input, conditioned on the obstacle’s planar position.

  • BarrierNet B(z,o)B(z,o): 11-D input, conditioned on fully random obstacle parameters.

Algorithm 2: Closed-form CBF safety correction
Input: Latent state $z_k$, obstacle $o$, decay rate $\alpha$, step $\Delta t$, barrier $B$
Output: Safe latent state $z_{\mathrm{safe}}$
// predictor: nominal goal-directed step
$z_{\mathrm{nom}} \leftarrow z_k-\operatorname{Adam}(\nabla(\mathcal{L}_{\mathrm{goal}}+\mathcal{L}_{\mathrm{prior}}))$;
$B_{\mathrm{target}} \leftarrow (1-\alpha\Delta t)B(z_k,o)$;
if $B(z_{\mathrm{nom}},o) \ge B_{\mathrm{target}}$ then
  $z_{\mathrm{safe}} \leftarrow z_{\mathrm{nom}}$;
else
  // corrector: project along the barrier gradient
  $g \leftarrow \nabla_z B(z_{\mathrm{nom}},o)$;
  $\kappa \leftarrow \min\left(1,\max\left(0,\frac{B_{\mathrm{target}}-B(z_{\mathrm{nom}},o)}{\lVert g\rVert^2+\epsilon}\right)\right)$;
  $z_{\mathrm{safe}} \leftarrow z_{\mathrm{nom}}+\kappa g$;
  Reset Adam state for $z$;
return $z_{\mathrm{safe}}$

Closed-Form Safety Correction

At planning time the obstacle term of (2) is removed and replaced by a predictor–corrector step (Algorithm 2). The predictor takes a nominal Adam step znomz_{\mathrm{nom}} on Lgoal+Lprior\mathcal{L}_{\mathrm{goal}}+\mathcal{L}_{\mathrm{prior}}. If B(znom,o)≥BtargetB(z_{\mathrm{nom}},o) \ge B_{\mathrm{target}}, the step is accepted; otherwise the corrector moves along the barrier gradient,

zsafe=znom+κ∇zB(znom,o).(6)z_{\mathrm{safe}}=z_{\mathrm{nom}}+\kappa\nabla_zB(z_{\mathrm{nom}},o). \tag*{(6)}

A first-order Taylor expansion around znomz_{\mathrm{nom}} gives

B(znom+κ∇zB)≈B(znom)+κ∥∇zB(znom)∥2,(7)B(z_{\mathrm{nom}}+\kappa\nabla_zB)\approx B(z_{\mathrm{nom}})+\kappa\lVert\nabla_zB(z_{\mathrm{nom}})\rVert^2, \tag*{(7)}

and setting the right-hand side equal to BtargetB_{\mathrm{target}} yields the smallest step along the gradient that restores the constraint to first order,

κ=max⁡(0,Btarget−B(znom,o)∥∇zB(znom,o)∥2+ϵ),(8)\kappa=\max\left(0,\frac{B_{\mathrm{target}}-B(z_{\mathrm{nom}},o)}{\lVert\nabla_zB(z_{\mathrm{nom}},o)\rVert^2+\epsilon}\right), \tag*{(8)}

which we clip to κ≤1\kappa\le1 to bound the correction magnitude. Optimizer state reset. When the correction overwrites zz, the Adam moment estimates no longer match the new iterate, and the accumulated momentum would push back against the projection in subsequent steps. We therefore reset Adam’s internal state for zz whenever κ>0\kappa>0.

Role of the Decay Rate

The decay rate α\alpha sets the safety–performance trade-off. As α→0\alpha\to0, condition (4) becomes B(zk+1)≥B(zk)B(z_{k+1})\ge B(z_k) and forbids any decrease of the barrier, producing an overly conservative planner that stalls near obstacles. As αΔt→1\alpha\Delta t\to1, the condition reduces to B(znom)≥0B(z_{\mathrm{nom}})\ge0, which permits large latent steps that violate the local linearity behind (8) and lead to collisions after decoding. Best performance is therefore expected at intermediate values, which we characterize empirically in Section VI.

Experimental Setup

Research Questions

The evaluation addresses three questions: (Q1) Does the closed-form CBF correction outperform a tuned binary classifier under identical planning conditions? (Q2) How does α\alpha mediate the safety–performance trade-off? (Q3) How does conditioning the barrier on increasingly general obstacle parameters affect generalization?

Environment, Data, and Evaluation Protocol

The workspace respects the joint limits of the Franka Panda (e.g., q1∈[−166∘,166∘]q_1 \in[-166^\circ,166^\circ]). The training corpus contains 100,000 free-space transitions and 100,000 obstacle collision states sampled uniformly within x∈[0.4,0.8]x \in[0.4,0.8] m and y∈[−0.5,0.5]y \in[-0.5,0.5] m. Every method is evaluated with the same two-stage protocol. In Stage 1, the planner produces a latent trajectory using its learned safety signal. In Stage 2, the decoded joint trajectory is validated in an independent geometric simulator (Robo3D) that models robot links as capsules and computes exact minimum distances to cylindrical obstacles. A plan succeeds only if the end effector ends within 10 mm of the target and no waypoint intersects an obstacle, so learned models never certify their own success.

Experimental Settings

Each of the three settings adds a degree of freedom to the obstacle parameterization:

  • S1, fixed obstacle. The obstacle is fixed at o=[0.5,0.0,0.75,0.1]o = [0.5, 0.0, 0.75, 0.1] and the target at e=[0.399,0.218,0.656]e = [0.399, 0.218, 0.656]. FixedBarrierNet is compared with a fixed-obstacle classifier of identical depth. To isolate the safety mechanism, GECO weight scheduling is disabled for both, with learning rate 0.03 and λprior=0.01\lambda_{\mathrm{prior}} = 0.01.

  • S2, multiple positions. MultiPosBarrierNet must handle four planar obstacle positions: (0.4,−0.2)(0.4,-0.2), (0.5,0.0)(0.5,0.0), (0.6,0.2)(0.6,0.2), and (0.5,0.3)(0.5,0.3).

  • S3, general planning. BarrierNet is evaluated on 500 randomized scenarios that match the conditions of the baseline experiment.

Ablations and Metrics

For every CBF variant, α\alpha is swept from 0.00 to 1.00 in increments of 0.05 (21 values). For the classifier, a 5×75 \times7 grid over λprior\lambda_{\mathrm{prior}} and λcollision\lambda_{\mathrm{collision}} (35 configurations) ensures that the CBF is compared with the best classifier configuration. Each configuration is evaluated on 500 scenarios. We report the success rate (target reached within 10 mm with no collision), the goal rate (target reached, ignoring collisions), and the collision-free (CF) rate (no collision, ignoring convergence). Separating goal and CF rates distinguishes failures caused by over-conservatism from failures caused by insufficient safety enforcement. For the CBF we also report the intervention rate, the fraction of optimization steps in which the corrector is applied, and the mean planning time per query. Hyperparameters are listed in Appendix A.

Success (%)Goal (%)CF (%)Time (ms)
70.579.279.8186.3

Table I. Reproduced LSPP baseline (1,000 scenarios)

Results

Baseline Reproduction

Evaluated on 1,000 random scenarios, the reproduced baseline achieves a 70.5% success rate (Table I). A further 9.3% of scenarios are collision-free but fail to converge, and 8.7% reach the goal but clip an obstacle. The bottleneck is gradient quality near obstacle boundaries: the classifier’s gradient is imperceptible far from an obstacle and dominant upon entry, which traps the optimizer in local minima.

Fixed Obstacle: CBF versus Classifier

Table II shows a non-monotonic success curve. At α=0\alpha= 0 the corrector intervenes in 90.8% of steps and nearly freezes the optimizer: the goal rate collapses to 56.4% and success to 46.8%. As α\alpha grows, the goal rate rises steadily to 99.0% at α=1\alpha= 1, while the CF rate falls as aggressive steps break the first-order approximation. The best trade-off occurs at α=0.20\alpha= 0.20, with 75.6% success (91.4% goal, 82.0% CF). Here the corrector is active in 37.9% of steps, often enough to enforce safety yet sparsely enough for the goal-directed optimizer to make progress.

Table III compares the best CBF settings with the best classifier settings. Across all 35 grid configurations, the classifier peaks at 63.6% (λc=5.0\lambda_c = 5.0, λprior=0.01\lambda_{\mathrm{prior}} = 0.01), so the CBF improves success by 12.0 percentage points under identical base conditions (Q1). The classifier reaches its best result only with heavy weighting that suppresses goal seeking (79.4% goal rate), whereas the predictor–corrector structure of the CBF avoids this gradient conflict.

Multi-Position Generalization

Conditioning the barrier on the obstacle position yields the best result overall: 83.8% success at α=0.10\alpha= 0.10 (93.6% goal, 89.8% CF), exceeding the fixed-obstacle barrier (Table II). We attribute this to an implicit regularization effect: interpolating across obstacle positions forces the network to learn a smoother barrier surface with a richer gradient field than fitting a single configuration does. Even modest spatial conditioning can therefore improve the learned barrier.

General Random-Obstacle Planning

The fully general BarrierNet B(z,o)B(z,o) degrades substantially, reaching at best 33.0% success at α=0.05\alpha= 0.05 (Table IV). We identify two causes (Q3):

Dimensionality. The 11-D input must represent a complex boundary over random radii and heights, and 15,000 transition scenarios are insufficient for this generalization. Similar challenges arise in offline safe RL [45] and conservative value estimation [43, 44].

α\alphaFixedBarrierNet B(z)B(z), S1 Success (%)Goal (%)CF (%)Interv. (%)Time (ms)MultiPosBarrierNet B(z,x,y)B(z,x,y), S2 Success (%)Goal (%)CF (%)
0.0046.856.488.290.8180.836.645.290.8
0.0574.486.287.067.0117.479.488.690.2
0.1074.488.285.052.399.683.893.689.8
0.1575.090.083.243.595.683.094.887.6
0.2075.691.482.037.994.881.695.085.2
0.2574.092.279.234.192.379.695.882.8
0.3071.092.275.831.786.879.696.082.4
1.0058.099.058.419.7117.056.698.457.4

Table II. Decay-rate ablation for the fixed-obstacle (S1) and multi-position (S2) barriers. Best success rates in bold.

MethodSuccess (%)Goal (%)CF (%)Time (ms)
CBF, α=0.20\alpha= 0.2075.691.482.094.8
CBF, α=0.15\alpha= 0.1575.090.083.295.6
CBF, α=0.10\alpha= 0.1074.488.285.099.6
Classifier, λc=5.0\lambda_c = 5.063.679.468.292.9
Classifier, λc=10.0\lambda_c = 10.063.278.270.0110.1
Classifier, λc=0.5\lambda_c = 0.558.891.060.4101.7

Table III. Best CBF versus best classifier configurations (S1).

α\alphaSuccess (%)Goal (%)CF (%)
0.0019.240.458.4
0.0533.069.439.8
0.10–1.00≈28.0–31.0up to 75.8≈30.0

Table IV. Decay-rate ablation for the general barrier B(z,o)B(z,o) (S3).

Covariate shift. During planning, the average pre-correction barrier value is −0.2786-0.2786. Because the transitions were generated by an obstacle-blind planner, the barrier routinely encounters deeply unsafe states (B<0B < 0) at deployment. There, the target (1−αΔt)B(zk)(1-\alpha\Delta t)B(z_k) is itself negative and the constraint no longer demands a meaningful correction: a CBF maintains forward invariance of the safe set but is not designed to recover from deep violations. This mirrors covariate shift in imitation learning [15, 56] and extrapolation error in offline RL [40, 42].

Computational Cost

The corrector evaluates the barrier gradient only at steps where the constraint is violated (37.9% of steps at α=0.20\alpha= 0.20), so the safety filter adds little overhead. In S1, the CBF planner averages 94.8 ms per query, on par with the best classifier (92.9 ms) and about half the 186.3 ms of the reproduced baseline configuration.

Limitations and Future Work

The main limitation is the covariate shift in the general random-obstacle task. A barrier trained on uncorrected trajectories cannot reliably enforce invariance inside an interactive optimization loop that drifts into unmapped regions of the 11-D input space, echoing distribution shift in offline RL [40, 41] and imitation learning [17]. In addition, the first-order correction assumes local linearity, but the VAE decoder is highly non-linear, so a step that is safe in latent space can occasionally decode into an unsafe motion in configuration space. These findings suggest three directions:

Iterative data aggregation. Generating transitions with the CBF-corrected planner itself and feeding them back into training, in the spirit of DAgger [56], directly targets the covariate shift; related ideas have succeeded in imitation [15, 17] and goal-conditioned learning [18].

Latent model predictive control. The current planner produces open-loop trajectories. A closed-loop MPC formulation [7] that operates sequentially in the latent space would let the barrier react to decoding errors at execution time.

Generative safety-aware planning. Diffusion-based planning [19] and feasibility-guided diffusion for offline safe RL [45] could combine expressive generative models with CBF constraints. Reachability-based flow policies [46], epigraph-guided flow matching for trading off safety and performance [47], and safety filters learned from offline data [48] are particularly relevant for integrating safety-aware generative planning with latent-space methods, while robust imitation through stochastic encoding [51] could mitigate the distribution shift identified here. HJ reachability [31, 32] and conformal-prediction-based neural CBFs [29] could further strengthen the formal safety guarantees.

Conclusion

We introduced a neural control barrier function that operates directly in the latent space of a generative manipulation planner. By decoupling obstacle avoidance from goal seeking and enforcing the barrier through a closed-form, first-order correction, the method removes the cliff-edge gradients of classifier-based safety. It achieves 75.6% and 83.8% success on fixed-obstacle and multi-position tasks, outperforming the best-tuned classifier by 12.0 percentage points at comparable planning cost, and the decay rate α\alpha provides a single, interpretable knob for the safety–performance trade-off, with the best range at [0.10,0.20][0.10, 0.20]. The general random-obstacle setting shows that barriers trained on obstacle-blind transitions cannot recover from the unsafe states visited at deployment. Together, these results establish the feasibility of CBF-based safety constraints in learned latent spaces and lay the groundwork for closed-loop, model-predictive formulations of latent-space manipulation planning.

Implementation Details

All models are implemented in PyTorch and trained on a single NVIDIA GPU; hyperparameters are listed in Table V. The barrier networks share the classifier’s architecture (four hidden layers of 2048 ELU units) but output an unbounded scalar, and their input dimension is 7, 9, or 11 depending on the variant. CBF planning uses a lower learning rate than the baseline (0.03 versus 0.15) because the correction relies on a first-order approximation that larger steps would violate, and a lower prior weight (0.01) because the correction itself limits the out-of-distribution drift that the prior term was designed to prevent. The baseline planning weights are the initial values of the GECO schedule and were varied in the classifier grid search.

ParameterValue
VAE training
Learning rate / batch / epochs10−410^{-4} / 4096 / 16,000
GECO ggoalg_{\mathrm{goal}}, glrg_{\mathrm{lr}}, gαg_{\alpha}0.0008, 0.005, 0.95
Baseline AM planning
Learning rate / max steps0.15 / 300
Initial λprior\lambda_{\mathrm{prior}} / λcollision\lambda_{\mathrm{collision}}0.7 / 0.5
Classifier temperature TT3.0
GECO αGECO\alpha_{\mathrm{GECO}}0.008
GECO τprior\tau_{\mathrm{prior}} / τobs\tau_{\mathrm{obs}}6.0 / 2.0
GECO moving averages (prior / obs.)0.8 / 0.8
CBF training
Learning rate10−410^{-4}, cosine to 10−610^{-6}
Batch size (label + transition)4096 + 4096
Margin γ\gamma1.0
Loss weights (λs\lambda_s, λu\lambda_u, λd\lambda_d)(1.0, 1.0, 1.0)
CBF planning
Learning rate / max steps0.03 / 300
Prior weight λprior\lambda_{\mathrm{prior}}0.01
Decay rate α\alpha / step Δt\Delta t0.20 / 1.0
Success threshold (all planners)10 mm

Table V. Hyperparameters

References

  1. [1]Q. Gallouédec, N. Cazin, E. Dellandréa, and L. Chen, “panda-gym: Open-Source Goal-Conditioned Environments for Robotic Learning,” 4th Robot Learning Workshop: Self-Supervised and Lifelong Learning at NeurIPS, 2021.arxiv.org/abs/2106.13687
  2. [2]Q. Gottesman, F. Johansson, M. Komorowski, A. Faisal, D. Sontag, F. Doshi-Velez, and L. A. Celi, “Guidelines for reinforcement learning in healthcare,” Nature Medicine, vol. 25, no. 1, pp. 16–18, 2019.DOI
  3. [3]Q. Nguyen and K. Sreenath, “Safety-critical control for dynamical bipedal walking with precise footstep placement,” IFAC-PapersOnLine, vol. 48, no. 27, pp. 147–154, 2015.
  4. [4]S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,” The International Journal of Robotics Research, vol. 30, no. 7, pp. 846–894, 2011.arxiv.org/abs/1105.1186
  5. [5]N. Ratliff, M. Zucker, J. A. Bagnell, and S. Srinivasa, “CHOMP: Gradient optimization techniques for efficient motion planning,” in IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2009, pp. 489–494.DOI
  6. [6]J. Schulman, J. Ho, A. X. Lee, I. Awwal, H. Bradlow, and P. Abbeel, “Finding locally optimal paths with sequential convex optimization,” in Robotics: Science and Systems (RSS), 2013.DOI
  7. [7]D. Mayne, J. Rawlings, C. Rao, and P. Scokaert, “Constrained model predictive control: Stability and optimality,” Automatica, vol. 36, no. 6, pp. 789–814, 2000.DOI
  8. [8]D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in International Conference on Learning Representations (ICLR), 2014.arxiv.org/abs/1312.6114
  9. [9]I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-VAE: Learning basic visual concepts with a constrained variational framework,” in International Conference on Learning Representations, 2017.
  10. [10]C.-M. Hung, F. Ebert, M. Mukadam, J. A. Bagnell, and S. Levine, “Reaching through latent space: From joint statistics to path planning in manipulation,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 5334–5341, 2022.DOI
  11. [11]B. Ichter and M. Pavone, “Robot motion planning in learned latent spaces,” in IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 2440–2446.arxiv.org/abs/1807.10366
  12. [12]A. D. Ames, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs with application to adaptive cruise control,” in 53rd IEEE Conference on Decision and Control. IEEE, 2014, pp. 6271–6278.DOI
  13. [13]A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control barrier functions: Theory and applications,” 2019 18th European Control Conference (ECC), pp. 3420–3431, 2019.arxiv.org/abs/1903.11199
  14. [14]C. Dawson, S. Gao, and C. Fan, “Safe control with learned certificates: A survey of neural Lyapunov, barrier, and contraction methods for robotics and control,” IEEE Transactions on Robotics, 2023.DOI
  15. [15]D. A. Pomerleau, “Efficient training of artificial neural networks for autonomous navigation,” Neural Computation, vol. 3, no. 1, pp. 88–97, 1991.DOI
  16. [16]A. Y. Ng and S. J. Russell, “Algorithms for inverse reinforcement learning,” in Proceedings of the Seventeenth International Conference on Machine Learning, 2000, pp. 663–670.
  17. [17]J. Ho and S. Ermon, “Generative adversarial imitation learning,” Advances in Neural Information Processing Systems, vol. 29, 2016.arxiv.org/abs/1606.03476
  18. [18]Y. Ding, C. Florensa, P. Abbeel, and M. Philipp, “Goal-conditioned imitation learning,” in Advances in Neural Information Processing Systems, vol. 32, 2019.arxiv.org/abs/1906.05838
  19. [19]M. Janner, Y. Du, J. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” in International Conference on Machine Learning. PMLR, 2022, pp. 9902–9915.arxiv.org/abs/2205.09991
  20. [20]A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2017.DOI
  21. [21]A. J. Barry, A. Majumdar, and R. Tedrake, “Safety verification of reactive controllers for UAV flight in cluttered environments using barrier certificates,” in 2012 IEEE International Conference on Robotics and Automation. IEEE, 2012, pp. 484–490.
  22. [22]M. Tayal, R. Singh, J. Keshavan, and S. Kolathaya, “Control barrier functions in dynamic UAVs for kinematic obstacle avoidance: a collision cone approach,” in 2024 American Conference (ACC). IEEE, 2024, pp. 3722–3727.arxiv.org/abs/2303.15871
  23. [23]C. Dawson, Z. Qin, S. Gao, and C. Fan, “Safe nonlinear control using robust neural Lyapunov-barrier functions,” in Conference on Robot Learning. PMLR, 2022, pp. 1724–1735.arxiv.org/abs/2109.06697
  24. [24]A. Robey, H. Hu, L. Lindemann, H. Zhang, D. V. Dimarogonas, S. Tu, and N. Matni, “Learning control barrier functions from expert demonstrations,” in Proceedings of the IEEE Conference on Decision and Control (CDC). IEEE, 2020, pp. 3710–3717.arxiv.org/abs/2004.03315
  25. [25]M. Srinivasan, A. Elbash, and S. Coogan, “Synthesis of control Lyapunov-Barrier functions with application to safety-critical control,” IEEE Control Systems Letters, vol. 4, no. 3, pp. 572–577, 2020.
  26. [26]H. Zhang, J. Wu, Y. Vorobeychik, and A. Clark, “Exact verification of ReLU neural control barrier functions,” in Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 5685–5705.arxiv.org/abs/2310.09360
  27. [27]A. Abate, D. Ahmed, A. Edwards, M. Giacobbe, and A. Peruffo, “FOS-SIL: A software tool for the formal synthesis of Lyapunov functions and barrier certificates using neural networks,” in Proceedings of the 24th International Conference on Hybrid Systems: Computation and Control, 2021, pp. 1–11.
  28. [28]O. So, Z. Serlin, M. Mann, J. Gonzales, K. Rutledge, N. Roy, and C. Fan, “How to train your neural control barrier function: Learning safety filters for complex input-constrained systems,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 11 532–11 539.
  29. [29]M. Tayal, A. Singh, P. Jagtap, and S. Kolathaya, “Cp-NCBF: A conformal prediction-based approach to synthesize verified neural control barrier functions,” arXiv preprint arXiv:2503.17395, 2025.arxiv.org/abs/2503.17395
  30. [30]M. Tayal, A. Singh, S. Kolathaya, and S. Bansal, “A physics-informed machine learning framework for safe and optimal control of autonomous systems,” in Forty-second International Conference on Machine Learning, 2025.
  31. [31]S. Bansal, M. Chen, S. Herbert, and C. J. Tomlin, “Hamilton-Jacobi reachability: A brief overview and recent advances,” in 2017 IEEE 56th Annual Conference on Decision and Control (CDC). IEEE, 2017, pp. 2242–2253.arxiv.org/abs/1709.07523
  32. [32]J. F. Fisac, N. F. Lugovoy, V. Rubies-Royo, S. Ghosh, and C. J. Tomlin, “Bridging Hamilton-Jacobi safety analysis and reinforcement learning,” in 2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 8550–8556.DOI
  33. [33]J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust control barrier-value functions for safety-critical control,” in 2021 60th IEEE Conference on Decision and Control (CDC). IEEE, 2021, pp. 6814–6821.arxiv.org/abs/2104.02808
  34. [34]J. García and F. Fernández, “A comprehensive survey on safe reinforcement learning,” Journal of Machine Learning Research, vol. 16, no. 1, pp. 1437–1480, 2015.DOI
  35. [35]W. Zhao, T. He, R. Chen, T. Wei, and C. Liu, “Safe reinforcement learning: A survey,” arXiv preprint arXiv:2302.03122, 2023.arxiv.org/abs/2302.03122
  36. [36]E. Altman, Constrained Markov Decision Processes. Routledge, 2021.DOI
  37. [37]J. Achiam, D. Held, A. Tamar, and P. Abbeel, “Constrained policy optimization,” in Proceedings of the 34th International Conference on Machine Learning. JMLR.org, 2017, pp. 22–31.arxiv.org/abs/1705.10528
  38. [38]M. Alsheikh, R. Bloem, R. Ehlers, B. Könighofer, S. Niekum, and U. Topcu, “Safe reinforcement learning via shielding,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, 2018.
  39. [39]R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. Cambridge, MA: The MIT Press, 1998.DOI
  40. [40]S. Levine, A. Kumar, G. Tucker, and J. Fu, “Offline reinforcement learning: Tutorial, review, and perspectives on open problems,” arXiv preprint arXiv:2005.01643, 2020.arxiv.org/abs/2005.01643
  41. [41]S. Fujimoto, D. Meger, and D. Precup, “Off-policy deep reinforcement learning without exploration,” in International Conference on Machine Learning. PMLR, 2019, pp. 2052–2062.arxiv.org/abs/1812.02900
  42. [42]A. Kumar, J. Fu, M. Soh, G. Tucker, and S. Levine, “Stabilizing off-policy Q-learning via bootstrapping error reduction,” Advances in Neural Information Processing Systems, vol. 32, 2019.arxiv.org/abs/1906.00949
  43. [43]A. Kumar, A. Zhou, G. Tucker, and S. Levine, “Conservative Q-learning for offline reinforcement learning,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 1179–1191.arxiv.org/abs/2006.04779
  44. [44]I. Kostrikov, A. Nair, and S. Levine, “Offline reinforcement learning with implicit Q-learning,” in International Conference on Learning Representations, 2022.arxiv.org/abs/2110.06169
  45. [45]Y. Zheng, J. Li, D. Yu, Y. Yang, S. E. Li, X. Zhan, and J. Liu, “Safe offline reinforcement learning with feasibility-guided diffusion model,” in International Conference on Learning Representations, 2024.arxiv.org/abs/2401.10700
  46. [46]M. Tayal, M. Tayal, and R. Prakash, “Safe flow Q-learning: Offline safe reinforcement learning with reachability-based flow policies,” arXiv preprint arXiv:2603.15136, 2026.arxiv.org/abs/2603.15136
  47. [47]M. Tayal, M. Tayal, and R. Prakash, “EpiFlow: Bridging safety and performance in offline reinforcement learning via epigraph-guided flow matching,” arXiv preprint arXiv:2602.08054, 2026.arxiv.org/abs/2602.08054
  48. [48]M. Tayal, M. Tayal, A. Singh, S. Kolathaya, and R. Prakash, “V-OCBF: Learning safety filters from offline data via value-guided offline control barrier functions,” arXiv preprint arXiv:2512.10822, 2025.arxiv.org/abs/2512.10822
  49. [49]R. K. Cosner, Y. Yue, and A. D. Ames, “End-to-end imitation learning with safety guarantees using control barrier functions,” in 2022 IEEE 61st Conference on Decision and Control (CDC). IEEE, 2022, pp. 5316–5322.arxiv.org/abs/2212.11365
  50. [50]F. Castañeda, H. Nishimura, R. T. McAllister, K. Sreenath, and A. Gaidon, “In-distribution barrier functions: Self-supervised policy filters that avoid out-of-distribution states,” in Proceedings of The 5th Annual Learning for Dynamics and Control Conference. PMLR, 2023, pp. 286–299.arxiv.org/abs/2301.12012
  51. [51]M. Tayal, M. Tayal, and R. Prakash, “RISE: Robust imitation through stochastic encoding,” arXiv preprint arXiv:2503.12243, 2025.arxiv.org/abs/2503.12243
  52. [52]Y. U. Ciftci, D. Chiu, Z. Feng, G. S. Sukhatme, and S. Bansal, “SAFE-GIL: SAFEty guided imitation learning for robotic systems,” arXiv preprint arXiv:2404.05249, 2024.arxiv.org/abs/2404.05249
  53. [53]Y. Wang, S. Z. Shan, R. Jiao, Z. Wang, W. Jin, Z. Yang, Z. Wang, C. Huang, and Q. Zhu, “Enforcing hard constraints with soft barriers: Safe reinforcement learning in unknown stochastic environments,” in International Conference on Machine Learning. PMLR, 2023, pp. 36 593–36 604.arxiv.org/abs/2209.15090
  54. [54]M. Tayal, A. Singh, P. Jagtap, and S. Kolathaya, “Semi-supervised safe visuomotor policy synthesis using barrier certificates,” arXiv preprint arXiv:2409.12616, 2024.arxiv.org/abs/2409.12616
  55. [55]D. J. Rezende and F. Viola, “Taming VAEs,” arXiv preprint arXiv:1810.00597, 2018.arxiv.org/abs/1810.00597
  56. [56]S. Ross, G. Gordon, and J. A. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in International Conference on Artificial Intelligence and Statistics (AISTATS), 2011.arxiv.org/abs/1011.0686

Paper details

vlnc:W_6e69fr2k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents