betaFeedback

WorksComputer Science

Elastra: A Simulation Framework for Elastic Bodies

Abstract

When robots interact with elastic surfaces such as mattresses and trampolines, changes in loading cause the surfaces to deform and rebound, altering subsequent contact locations and forces. We present Elastra, a simulation framework for modeling this interaction and supporting the training and evaluation of robot control policies. Elastra represents each surface using elements that move only vertically, with elastic connections and damping to model deformation and recovery. Contact forces drive surface motion, and the updated surface geometry is used to compute subsequent contact forces. We systematically validate our simulations against an independent implementation and a real mattress load curve. We further use Elastra to train the Unitree G1 humanoid to stand up on mattresses and trampolines. Using the pretrained HoST policy and ProtoMotions motion tracker as original controllers, we train residual policies through reinforcement learning to add corrections to their joint targets. Adding these residual policies increases the success rate with HoST from 44.1% to 72.5%, and from 55% to 89% on the trampoline. With ProtoMotions, the success rate increases from 19.0% to 45.0% for initial poses not encountered during training. The code and video are available on https://github.com/yichg/elastra.

Introduction

To use the motion skills of a robot in different environments, one has to consider how changed contact conditions affect the execution of a motion. Mattresses, sofa cushions and trampolines sink where they are pressed and recover when the load moves on, so every push changes where the next contact lands and what force it meets. Robot learning environments usually treat the ground as rigid. This paper builds Elastra, a simulation framework in which mattresses and trampolines deform under the robot and the deformed surface enters the contact computation. As the demonstration task we use standing up after a fall (Figure 1), which loads and unloads the surface with several body parts in turn. HoST [1] and HumanUP [2] have demonstrated humanoid standing up from many fallen postures and on varied terrain, and Singh et al. [3] achieved humanoid walking on a mattress and soft foam. This paper instead simulates the support itself as a body that moves under the robot.

The Unitree G1 stands up from a prone pose on a mattress simulated with Elastra

Figure 1. The Unitree G1 stands up from a prone pose on a mattress simulated with Elastra, with a residual controller adapted from HoST. The mattress sinks under each contact and recovers as the load moves. Trampoline get-up is reported in Figure 3.

The contributions are (1) Elastra, a mattress and trampoline simulation framework for robot learning that incorporates local deformation, neighboring-region motion, and post-unloading recovery into the contact computation, (2) a systematic validation of Elastra against (i) an independent implementation and (ii) a real mattress load curve, and (3) a demonstration in which residual reinforcement learning on top of a pretrained standing policy enables the Unitree G1 to stand up on both mattresses and trampolines, evaluated across different stiffness levels and on fallen postures unseen during training.

Related Work

Simulation of deformable supports. The outcome of a contact simulation depends on the shapes, materials and contact parameters of the bodies and on the numerical method that solves the equations of motion. In robotics simulators, the PhysX deformable bodies of Isaac Sim [4], Warp [5], the flex bodies of MuJoCo [6, 7] and SOFA [8] represent soft volumes and cloth, and benchmarks such as SoftGym [9] and PlasticineLab [10] use them for manipulation. Elastra is built inside MuJoCo and uses its flex bodies only for collision geometry, not for their elasticity. A flex body is a volumetric finite element model with linear elements, so the elements must be small enough to follow the deformation under each contact. In the task of standing up on a mattress, the robot presses on the mattress with feet, hands and knees, and each contact loads a patch about 0.1 m across. The compression under such a patch fades within about one patch width to the side and in depth, so the elements must be smaller than the patch in length, width and depth. On the mattress we consider, elements of this size amount to 10410^4 degrees of freedom. Elastra avoids this cost. Each cell of Elastra keeps one vertical degree of freedom, and a bending term carries the lateral spreading of a load, such that the state of the surface is the vector of cell heights and 380 degrees of freedom describe the same mattress. Every restoring and damping force follows an explicit physical law given in § 3, so the material law stays open to calibration by a real mattress. Together, the low cost and the observable surface state make Elastra friendly to reinforcement learning, and the response calibrated against a real mattress prepares the transfer to hardware. Sun et al. [11] describe such calibrated values as effective parameters that absorb the details omitted by a simulator and reproduce real behavior.

Robots on soft ground. Foam, soft pads, and similar ground deform under load, and a robot moving on them has to adapt to changes in foot placement and contact force. Singh et al. [3] varied the shape and the softness of the ground in training, achieved walking on a mattress and soft cushion foam with a real HRP-5P, and studied aperiodic walking with a variable step frequency. This paper computes, with movable cells and their connections, the surface motion caused by any body part that loads and unloads the support; standing-up exercises with hands, knees, torso, and feet in turn.

Standing up. A basic choice in the control of standing up is how to obtain the sequence of motions from lying to standing. Kanehiro et al. [12] searched a graph of postures to connect the current posture to an existing recovery motion. In physics based character animation, DeepMimic [13] combines the imitation of a reference motion with a task reward and learns a feedback policy that executes the motion from the current state. Motion trackers trained in frameworks such as ProtoMotions [14] bring this approach to robots. Learned standing control has also progressed on real humanoid robots. For full size humanoids, HoST [1] does not rely on a predefined standing trajectory. HumanUP [2] separates the discovery of the motion from the training of a deployable controller. These studies treat the ground as given, and our paper builds the simulation of supports that deform and recover under the robot, and uses standing up as a demonstration.

Residual reinforcement learning. Residual reinforcement learning [15, 16] adds a learned correction to the output of an existing controller, so that the task is completed with the original motion ability while the motion is adjusted from the new interaction. On legged and humanoid robots, reinforcement learning has likewise been combined with an existing controller [17, 18, 19, 20].

Mattress and Trampoline Models

We discretize the mattress and the trampoline in MuJoCo [6] into surface cells with mass, keep only the vertical translation of each cell, and describe deformation and recovery under load by elastic connections and damping. Let the surface have nsn_s movable nodes and let xi=ziref−zix_i = z_i^{\mathrm{ref}} - z_i be the downward displacement of node ii, where zirefz_i^{\mathrm{ref}} is its reference height and ziz_i its current height. Collecting the displacements of the movable nodes in a vector x∈Rnsx \in\mathbb{R}^{n_s}, the nodes move according to

Msx¨+Cs(α)x˙+fel(x;α)=fcontact+fg.(1)M_s \ddot{x} + C_s(\alpha)\dot{x} + f_{\mathrm{el}}(x;\alpha) = f_{\mathrm{contact}} + f_g. \tag*{(1)}

Here Ms=diag⁡(m1,…,mns)M_s = \operatorname{diag}(m_1,\ldots,m_{n_s}) is the node mass matrix, Cs(α)x˙C_s(\alpha)\dot{x} the damping force and fel(x;α)f_{\mathrm{el}}(x;\alpha) the elastic restoring force, and α>0\alpha> 0 is a dimensionless stiffness scale factor. fcontactf_{\mathrm{contact}} is the vertical component of the contact forces of the robot on the nodes and fgf_g the weight of the nodes.

Mattress. The mattress is a rectangular grid of blocks above a rigid floor (Figure 2, left). Each block resists local compression through a vertical spring and damper, and for every interior cell a bending term resists the curvature of the surface at that cell. The mattress moves according to

Figure 2. The two surface models of Elastra. Top: a steel ball resting on the simulated mattress (left) and trampoline (right); the dashed lines mark the sections A–A and B–B drawn below. Bottom: mattress cells rest on local springs and dampers and are coupled to their neighbors through a bending term; trampoline nodes carry the load through the membrane connections to a fixed rim. The renders use softer parameters than the experiments so that the deformation is visible.

mx¨+cx˙+(kI+bLTL)x=fcontact+mg1,(2)m\ddot{x} + c\dot{x} + (kI + bL^{\mathsf{T}}L)x = f_{\mathrm{contact}} + mg\mathbf{1}, \tag*{(2)}

where L\mathrm{L} is the five point Laplacian over the interior cells, (Lx)i=xN+xS+xE+xW−4xi(\mathrm{L}x)_i = x_N + x_S + x_E + x_W - 4x_i, kk the cell stiffness, bb the bending stiffness, cc the cell damping and mm the cell mass, so that the elastic energy is k∣x∣2/2+b∣Lx∣2/2k|x|^2/2 + b|\mathrm{L}x|^2/2. The bending term spreads a local load over the neighboring cells; because every row of L\mathrm{L} sums to zero, the weight of the cells sags the unloaded bed uniformly by mg/kmg/k and the surface the robot lands on is flat. There is no separate connection between pairs of cells and no nonlinear compression term. We change the stiffness of the mattress through the single factor α\alpha,

k(α)=αk1,b(α)=αb1,c(α)=α c1,(3)k(\alpha) = \alpha k_1,\qquad b(\alpha) = \alpha b_1,\qquad c(\alpha) = \sqrt{\alpha}\,c_1, \tag*{(3)}

where k1k_1, b1b_1 and c1c_1 are the reference values, so that the damping ratio of a cell and the shape of the dent under a load are the same at every stiffness and only its depth changes. Geometry, mass and contact parameters stay fixed when the stiffness changes.

Trampoline. The trampoline is a circular network of nodes with a fixed rim, and its interior nodes translate vertically (Figure 2, right). Neighboring nodes transfer load through connections within the membrane, and no interior node has its own spring to the ground. For an interior node ii the elastic term of (1) is

fel,i(x;α)=∑j∈N(i)[klin(α)(xi−xj)+kcub(α)(xi−xj)3](4)f_{\mathrm{el},i}(x;\alpha) = \sum_{j\in\mathcal{N}(i)} \left[k_{\mathrm{lin}}(\alpha)(x_i-x_j) + k_{\mathrm{cub}}(\alpha)(x_i-x_j)^3\right] \tag*{(4)}

where N(i)\mathcal{N}(i) is the set of the four neighbors of node ii, and klink_{\mathrm{lin}} and kcubk_{\mathrm{cub}} are the linear and cubic coefficients of the restoring force. A node on the fixed rim has xj=0x_j = 0. Stiffness scales with α\alpha and the damping with α\sqrt{\alpha}. Collisions between the robot and the trampoline are handled by a two dimensional flex mesh of MuJoCo that covers the node network [7]. Its vertices move with the node bodies and the contact forces are distributed to the nodes involved; the elasticity of the flex mesh itself is not used, and the restoring forces between nodes follow (4).

A Demonstration with Residual Reinforcement Learning

Two pretrained controllers provide the original motion, and residual reinforcement learning trains an additional control term. During training the original controller is frozen. Its joint targets plus the residual output go to the original proportional derivative controller, which turns them into torque (Figure 5).

Figure 5. Demonstration controller. The pretrained network keeps its weights, a residual policy adds a bounded correction to its joint targets, and the sum goes to the original proportional derivative controller. The residual policy reads the robot state and the output of the original controller. The HoST residual also reads information about the surface.

Original controllers and residual parameterization. Our first controller is the HoST [1] standing policy. It generates joint position increments from the robot state, and the released prone or supine policy is selected from the initial posture. With dd controlled joints, current joint angles qt∈Rdq_t \in\mathbb{R}^d and HoST output ato∈Rda_t^o \in\mathbb{R}^d, the original joint target is q0,tdes=qt+s0atoq_{0,t}^{\mathrm{des}} = q_t + s_0a_t^o, where s0>0s_0 > 0 is the action scale of the original policy. The second controller is the ProtoMotions [14] motion tracker, whose deployment interface produces the joint target q0,tdesq_{0,t}^{\mathrm{des}} from the robot state and a reference motion. Let oto_t be the observation the residual policy receives at control step tt. During training the policy is a diagonal Gaussian whose output is bounded by a hyperbolic tangent,

atδ∼N(μθ(ot),diag⁡(σ2)),δt=stanh⁡(atδ),qtdes=q0,tdes+δt,(5)\begin{aligned} a_t^\delta&\sim\mathcal{N}\left(\mu_\theta(o_t),\operatorname{diag}(\sigma^2)\right),\\ \delta_t &= s\tanh(a_t^\delta),\\ q_t^{\mathrm{des}} &= q_{0,t}^{\mathrm{des}} + \delta_t, \tag*{(5)} \end{aligned}

where atoa_t^o is the action sampled from the policy, δt\delta_t is the residual correction added to the joint targets, and qtdesq_t^{\mathrm{des}} is the joint target sent to the joint controller. The transform acts componentwise, and s>0s > 0 is the bound on the correction of a joint target.

Observation and task objective. The observation of the residual policy contains the robot state and the output of the original controller. For HoST it also contains information about the surface, and the second training stage adds the position of the robot in the task area (Table 2). Let ztpelvisz^{\mathrm{pelvis}}_t be the pelvis height in the world frame, ztsupz^{\mathrm{sup}}_t the height of the top of the mattress cell or trampoline node directly below the lowest point of the collision geometry of the robot (ztsup=0z^{\mathrm{sup}}_t = 0 on rigid ground), RWB(t)R_{WB}(t) the rotation matrix from base coordinates to world coordinates and ez=(0,0,1)⊤\mathbf{e}_z = (0,0,1)^\top. The uprightness of the base and the standing indicator are

InputHoSTTrackerRepresentation
Joint positions2329joint angles
Joint velocities2329factor 0.1
Base orientation44quaternion
Base angular velocity33base frame, factor 0.25
Original controller output2329HoST increment, tracker target
Torso height, torso uprightness, pelvis height33three scalars
Surface type3–one hot encoding
Stiffness factor1–log⁡2α/6\log_2 \alpha/6, rigid ground 1
Maximum surface deflection1–factor 10
Robot position7–root xyx y, root velocity with factor 0.1, distances to the task area boundary; second training stage only
Progress of the trial11steps done over steps planned
Total9298

Table 2. Inputs of the residual policy. A factor is a fixed scaling applied before the input enters the network.

ut=ez⊤RWB(t)ez,It=1{ztpelvis−ztsup≥zmin⁡, ut≥umin⁡},(6)u_t = \mathbf{e}_z^\top R_{WB}(t)\mathbf{e}_z, \qquad I_t = 1\{z^{\mathrm{pelvis}}_t-z^{\mathrm{sup}}_t \ge z_{\min},\ u_t \ge u_{\min}\}, \tag*{(6)}

where zmin⁡z_{\min} and umin⁡u_{\min} are the height and uprightness thresholds. The pelvis is measured against the deformed support because on a soft mattress a robot standing upright sinks so far that its pelvis stays below zmin⁡z_{\min} in the world frame. A trial succeeds when the standing condition holds for TholdT_{\mathrm{hold}} control steps and the robot never leaves the task area. The reward at each control step is

rt=ht+wuuttorso+It+wJJt−wOOt,(7)r_t = h_t + w_u u^{\mathrm{torso}}_t + I_t + w_J J_t - w_O O_t, \tag*{(7)}

where ht∈[0,1]h_t \in[0,1] is the progress of the torso height from its initial value to the reference stand and uttorso∈[0,1]u^{\mathrm{torso}}_t \in[0,1] the uprightness of the torso, JtJ_t is one only at the step at which the trial first completes TholdT_{\mathrm{hold}} standing steps and zero otherwise, and OtO_t is one when the robot leaves the allowed area, which also ends the training trial. We train the residual policy with proximal policy optimization (PPO) [21].

Evaluation Setup

Table 1 lists the parameters of the two surfaces, Table 2 the inputs of the residual policy, and Table 3 the controller, task, and training parameters. For the mattress, the reference bed at α=1\alpha= 1 is a soft foam. The cell stiffness k1=105.30k_1 = 105.30 N/m and the bending stiffness b1=50.72b_1 = 50.72 N/m are fitted to the force against indentation of two rigid flat discs of 356 mm and 100 mm diameter pressed into a finely discretized continuum foam model in Warp [5] with 6.25 mm elements, with a pooled RMS error of 1.7%, and the cell damping c1=0.876c_1 = 0.876 N s/m gives a damping ratio of 0.25. For the trampoline, the reference at α=1\alpha= 1 is derived from force versus displacement data digitized from Park et al. [22].

HoST initial states follow the perturbation protocol of [1], prone and supine, with 16 training and 48 test states per posture drawn independently, so each surface has 96 paired tests. For ProtoMotions, the residual trains on 39 lying poses from the pose pool of HumanUP [2]. The 40 test poses are passive falls of the robot with its motors off on rigid ground from a perturbed standing pose. Both controllers are evaluated on rigid ground, seven mattresses and two trampolines.

We implement the mattress and the trampoline a second time in Warp [5], which shares no code with MuJoCo. This second implementation is separate from the continuum foam model used for the calibration. We compare the dynamic response after a steel ball is released at the surface center and its static deflection under different stiffness conditions. When comparing the static deflection, we also use the static solution of the same spring law from § 3 as a third reference, labeled analytic estimate in Figure 6. Vlaović et al. [23] pressed the loading pad of EN 1957 [24] into a polyurethane mattress and recorded the load against indentation. We fit the material scaling parameter to the shallow part of the curve and predict the load at the held-out indentation depths to further validate the physical accuracy of our simulation.

Figure 6. Steel ball on the simulated mattress (left) and trampoline (right). Static deflection versus stiffness in MuJoCo Elastra, in an independent Warp implementation, and from the analytic estimate.

Results

Systematic validation of the surface models. As shown in Figure 6, the static deflections in MuJoCo Elastra and Warp Elastra agree within 0.3 mm on the trampoline at all 6 stiffness levels and within 3 mm on the 4 comparable mattress levels. Both implementations agree with the analytic estimates. Moreover, MuJoCo Elastra and Warp Elastra produce closely matching deflection curves during the initial sinking, rebound, and decaying oscillations. The compression law also predicts the held-out loads of the real mattress [23] with relative errors from 5.9% to 10.8%.

HoST on mattresses and trampolines. A residual policy recovers get-up on both the mattress and the trampoline (Figure 3). On rigid ground, both the original controller and the residual succeed in every trial. On the three stiffer mattresses, the residual succeeds in 95 or 96 of 96 trials, and on the mattresses at α=8\alpha= 8 and α=4\alpha= 4 it raises success from 25% to 94% and from 6.3% to 45%. On the trampoline, the residual raises success from 55% to 89% at α=1\alpha= 1 and from 50% to 88% at α=2\alpha= 2.

Figure 3. HoST, 96 trials per surface. (a) Success rate of the original controller and of the residual policy per surface. (b) Outcomes over all the 960 trials in three classes. The robot never reaches the standing thresholds, reaches them but either loses them before 1 s or leaves the task area, or succeeds.

ProtoMotions on postures unseen in training. The residual generalizes to fallen postures that were not used in training. Figure 4 shows the ProtoMotions tracker on those postures, 40 trials on each mattress, trampoline and rigid ground, 400 trials in all. Success rises from 19.0% to 45.0%. The largest gains are on the two trampolines, from 10% and 15% to 52.5% and 60%.

Figure 4. ProtoMotions tracker on postures unseen in training, 40 trials per surface. (a) Success rate of the original tracker and of the residual policy per surface. (b) Outcomes over all the 400 trials in the three classes of Figure 3.

Conclusion and Future Work

We built Elastra, simulated mattress and trampoline environments for robot learning, validated them in four ways, and demonstrated them with humanoid get-up by adapting frozen pretrained controllers with a small residual policy. All robot experiments are simulated. The surfaces are level and uniform, with the same stiffness and damping in every cell, whereas real supports vary from place to place, so nonuniform media are a natural extension. The next steps are hardware experiments, nonuniform supports, other robots, and tasks such as a quadruped walking across a bed.

Acknowledgment

Generative language models assisted in preparing this manuscript. GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 were used to draft and revise prose in Sections 1 to 7 and to help prepare tables and figure captions, and to write the simulation and training code of Sections 3 to 6. All claims, configurations and results were verified by the authors, who take full responsibility for the content of this paper.

Parameters and Additional Figures

MattressTrampoline
Grid and size20×19 blocks of 0.10×0.10×0.05 m covering 2.0 × 1.9 m25 × 25 nodes at 0.1 m spacing, circular region of radius 1.2 m
Vertical travel0.247 m, rigid floor 0.26 m below the reference top0.43 m
Elastic parametersk1=105.30k_1 = 105.30 N/m, b1=50.72b_1 = 50.72 N/m, 306 bending terms, both ∝α\propto\alphaklin=3025.79αk_{\mathrm{lin}} = 3025.79\alpha N/m, kcub=10−9αk_{\mathrm{cub}} = 10^{-9}\alpha N/m3^{3}
Dampingc1=0.876c_1 = 0.876 N s/m, damping ratio 0.25, scales with α\sqrt{\alpha}scales with α\sqrt{\alpha}
Inertial mass0.02915 kg per block, about 11.1 kg in allnode mass fixed
Contact geometryrigid blocks moving with the cellsflex mesh moving with the nodes, collision radius 2 mm

Table 1. Surface parameters used for training and evaluation with the robot.

HoST23 actuated joints, action scale s0=0.25s_0 = 0.25, observation history of six control steps, trial 5.02 s, correction bound 0.33 rad (1.32 action units)
ProtoMotions29 actuated joints, clips of 425, 361 and 497 control steps (8.50, 7.22, 9.94 s), heading aligned at the start, correction bound 0.85 rad
Time stepssimulation step 0.625 ms, control period 20 ms (32 substeps)
Standing conditionpelvis height above the support below the lowest robot point ≥0.7\ge0.7 m, base uprightness ≥0.9\ge0.9, held 50 control steps (1 s)
Reward weightswu=0.5w_u = 0.5, wJ=20w_J = 20, wO=10w_O = 10
Networkspolicy and value, two hidden layers of 256 and 128, tanh, orthogonal init gain 2\sqrt{2}, policy output gain 0.01, initial log std −1-1
PPO256 environments ×\times 64 steps per update, 4 passes ×\times 4 minibatches, 1000 updates for HoST in two stages of 500 and 500 updates for ProtoMotions (16,384,000 and 8,192,000 steps), γ=0.995\gamma= 0.995, GAE 0.95, clip 0.2, lr 3×10−43 \times10^{-4}, value coef 0.5, entropy 0, grad norm 1.0, target KL 0.05
Training surfacesboth controllers: rigid ground, mattress α∈{8,16,32,64}\alpha\in\{8, 16, 32, 64\} and trampoline α∈{1,2}\alpha\in\{1, 2\}, mixed; second stage of HoST: mattress α∈{1,2,4,8}\alpha\in\{1, 2, 4, 8\} instead
Evaluation surfacesboth controllers on rigid ground, the mattress at α∈{1,2,4,8,16,32,64}\alpha\in\{1, 2, 4, 8, 16, 32, 64\} and the trampoline at α=1\alpha= 1 and 2

Table 3. Controller, task and training parameters.

References

  1. [1]T. Huang et al., “Learning humanoid standing-up control across diverse postures,” arXiv:2502.08378, 2025.arxiv.org/abs/2502.08378
  2. [2]X. He, R. Dong, Z. Chen, and S. Gupta, “Learning getting-up policies for real-world humanoid robots,” arXiv:2502.12152, 2025.arxiv.org/abs/2502.12152
  3. [3]R. P. Singh, M. Morisawa, M. Benallegue, Z. Xie, and F. Kanehiro, “Robust humanoid walking on compliant and uneven terrain with deep reinforcement learning,” arXiv:2504.13619, 2025, published at Humanoids 2024.DOI
  4. [4]NVIDIA, “Omni Physics Documentation 107.3, Deformable Bodies (Beta),” https://docs.omniverse.nvidia.com/kit/docs/omni_physics/107.3/dev_guide/deformables_beta/deformable_beta.html, 2025.
  5. [5]M. Macklin, “Warp: A high-performance Python framework for GPU simulation and graphics,” NVIDIA GPU Technology Conference (GTC), 2022.
  6. [6]E. Todorov, T. Erez, and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2012, pp. 5026–5033.DOI
  7. [7]Google DeepMind, “MuJoCo 3.3.3 documentation, modeling chapter,” https://mujoco.readthedocs.io/en/3.3.3/modeling.html, 2025.
  8. [8]F. Faure et al., “SOFA: A multi-model framework for interactive physical simulation,” in Soft Tissue Biomechanical Modeling for Computer Assisted Surgery. Springer, 2012, pp. 283–321.DOI
  9. [9]X. Lin, Y. Wang, J. Olkin, and D. Held, “SoftGym: Benchmarking deep reinforcement learning for deformable object manipulation,” arXiv:2011.07215, 2021.arxiv.org/abs/2011.07215
  10. [10]Z. Huang et al., “PlasticineLab: A soft-body manipulation benchmark with differentiable physics,” arXiv:2104.03311, 2021.arxiv.org/abs/2104.03311
  11. [11]Y. Sun, J. Guo, X. Ren, C. Yi, and H. Yang, “Effective parameters, real behavior: Renormalization for robotics – from infinite electron mass to sim-to-real gap,” arXiv:2607.24079, 2026.arxiv.org/abs/2607.24079
  12. [12]F. Kanehiro, K. Fujiwara, H. Hirukawa, S. Nakaoka, and M. Morisawa, “Getting up motion planning using Mahalanobis distance,” in IEEE International Conference on Robotics and Automation (ICRA), 2007, pp. 2540–2545.DOI
  13. [13]X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “DeepMimic: Example-guided deep reinforcement learning of physics-based character skills,” arXiv:1804.02717, 2018.
  14. [14]C. Tessler et al., “ProtoMotions3: An open-source framework for humanoid simulation and control,” https://github.com/NVlabs/ProtoMotions, 2025.
  15. [15]T. Johannink et al., “Residual reinforcement learning for robot control,” in IEEE International Conference on Robotics and Automation (ICRA), 2019, pp. 6023–6029, arXiv:1812.03201.arxiv.org/abs/1812.03201
  16. [16]T. Silver, K. Allen, J. Tenenbaum, and L. Kaelbling, “Residual policy learning,” arXiv:1812.06298, 2019.arxiv.org/abs/1812.06298
  17. [17]Y. Chen and Q. Nguyen, “Learning agile locomotion and adaptive behaviors via RL-augmented MPC,” arXiv:2310.09442, 2023.arxiv.org/abs/2310.09442
  18. [18]J. Kamohara, F. Wu, C. Wamorkar, S. Hutchinson, and Y. Zhao, “RL-augmented adaptive model predictive control for bipedal locomotion over challenging terrain,” arXiv:2509.18466, 2025.arxiv.org/abs/2509.18466
  19. [19]S. H. Jeon, H. J. Lee, S. Hong, and S. Kim, “Residual MPC: Blending reinforcement learning with GPU-parallelized model predictive control,” arXiv:2510.12717, 2025.arxiv.org/abs/2510.12717
  20. [20]J. Li, L. Wu, S. A. Esteban, L. Yang, J. Drgoňa, and A. D. Ames, “Accelerating and scaling MPC-guided reinforcement learning for humanoid locomotion and manipulation,” arXiv:2606.05687, 2026.arxiv.org/abs/2606.05687
  21. [21]J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv:1707.06347, 2017.arxiv.org/abs/1707.06347
  22. [22]G. Park, S.-H. Choi, C.-H. Kim, M. Y. Kim, and S. Lee, “Trampoline stiffness estimation by using robotic system for quantitative evaluation of jumping exercises,” Sensors, vol. 23, no. 24, p. 9645, 2023.DOI
  23. [23]Z. Vlaović, N. Klarić, and D. Domljan, “Investigating the impact of long-term use on mattress firmness and sleep quality—preliminary results,” Applied Sciences, vol. 14, no. 21, p. 10016, 2024.DOI
  24. [24]European Committee for Standardization (CEN), “EN 1957:2012, Furniture. Beds and mattresses. Test methods for the determination of functional characteristics and assessment criteria,” 2012.

Paper details

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

views · 6

downloads · 0

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

  • Computer Science

GPT-5.6 Sol (bare), Claude Opus 5 (bare), Grok 4.6 (bare)

View statement
Looking for related papers…
Contents