Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale

Chengkun Li; Cheryl Wang; Bianca Ziliotto; Merkourios Simos; Jozsef Kovecses; Guillaume Durandau; Alexander Mathis

cs.ROMar 26, 2026

Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale

Authors: Chengkun Li, Cheryl Wang, Bianca Ziliotto, Merkourios Simos, Jozsef Kovecses, Guillaume Durandau, Alexander Mathis

Organizations: EPFL, Switzerland · McGill University, Canada · Feil & Oberfeld CRIR-Jewish Rehabilitation Hospital Research Center, Canada · NVIDIA

Abstract

Learning motor control for muscle-driven musculoskeletal models is hindered by the computational cost of biomechanically accurate simulation and the scarcity of validated, open full-body models. Here we present MuscleMimic, an open-source framework for scalable motion imitation learning with physiologically realistic, muscle-actuated humanoids. MuscleMimic provides two validated musculoskeletal embodiments - a fixed-root upper-body model (126 muscles) for bimanual manipulation and a full-body model (416 muscles) for locomotion - together with a retargeting pipeline that maps SMPL-format motion capture data onto musculoskeletal structures while preserving kinematic and dynamic consistency. Leveraging massively parallel GPU simulation, the framework achieves order-of-magnitude training speedups over prior CPU-based approaches while maintaining comprehensive collision handling, enabling a single generalist policy to be trained on hundreds of diverse motions within days. The resulting policy faithfully reproduces a broad repertoire of human movements under full muscular control and can be fine-tuned to novel motions within hours. Biomechanical validation against experimental walking and running data demonstrates strong agreement in joint kinematics (mean correlation r = 0.90), while muscle activation analysis reveals both the promise and fundamental challenges of achieving physiological fidelity through kinematic imitation alone. By lowering the computational and data barriers to musculoskeletal simulation, MuscleMimic enables systematic model validation across diverse dynamic movements and broader participation in neuromuscular control research. Code, models, checkpoints, and retargeted datasets are available at: https://github.com/amathislab/musclemimic

Figures & tables

Model	Type	Joints	Muscles	DoFs	Action dim.	Focus
MyoBimanualArm	Fixed-base	76 (36 ∗ )	126 (64 ∗ )	54 (14 ∗ )	126 (64 ∗ )	Upper-body manipulation
MyoFullBody	Free-root	123 (83 ∗ )	416 (354 ∗ )	72 (32 ∗ )	416 (354 ∗ )	Whole body motion

Table 1: Overview of the two musculoskeletal learning embodiments in MuscleMimic. Both models support enabling and disabling finger muscles to facilitate faster convergence when fine finger control is not required ( ∗ denotes configurations with finger muscles disabled). Joints denote articulated connections, while DoFs (degrees of freedom) correspond to independently controllable joint coordinates. Different from traditional robotics systems, certain MSK model joints are dependent on each other, resulting in fewer DoFs than joints.

Figure 1: Visualization of the MyoBimanualArm model and MyoFullBody model, viewed from (A) front, (B) back, and (C) side on first and second row, respectively.

Figure 2: Total system throughput (Raw training Steps Per Second) as the number of parallel environments (

n

) scales from

16

8192

(the rest of the hyperparameters stay the same). Evaluated on an Intel Xeon Platinum 8570 CPU and a single NVIDIA H100 80GB GPU. Training was evaluated with a fixed number of mini-batches of 32 and 50 steps per rollout. With

8192

environments, the throughput increases by around

7800\%

Figure 3: Effect of gradient epochs (

E

) on training stability. We compare

E=1

(truly on-policy),

E=3

, and

E=10

(aggressive sample reuse). (A) Early training (first 30M steps): higher

E

accelerates initial learning due to more gradient updates per sample. (B) Full training trajectory:

E=1

achieves superior asymptotic performance while

E=3

and

E=10

plateau or collapse. (C) KL divergence between current and data-generating policy distributions (log scale); with the same amount of clipping,

E>1

exhibits catastrophic distribution shift with spikes exceeding

10^{10}

, whereas

E=1

remains stable below

10^{-1}

Figure 4: Effect of minibatch size on training dynamics. We compare minibatch sizes of 32, 64, and 128. (A) Performance: larger batch sizes achieve higher asymptotic rewards. (B) Exploration stability: smaller batches cause the policy standard deviation to overshoot, while larger batches maintain stable convergence near the initialization. (C) Policy update magnitude (log scale): larger batches yield lower KL divergence throughout training, indicating more conservative and stable policy updates.

Figure 5: Motion snapshots from pre-trained MyoBimanualArm policies (fingers disabled). From top to bottom: lifting objects, throwing a ball, waving, and pouring then placing water.

Figure 6: Motion snapshots from pre-trained MyoFullBody policies. From top to bottom: walking, running, turning in a circle, dancing, vertical jumping, and kick twist.

Metric	MyoFullBody		MyoBimanualArm
Metric	Training	Testing	Training	Testing
Success rate $\uparrow$	$95.51\pm 0.42\%$	$92.62\pm 0.01\%$	$98.64\pm 0.03\%$	$99.46\pm 0.02\%$
Frame Coverage $\uparrow$	$97.60\pm 0.31\%$	$96.04\pm 0.45\%$	$98.17\pm 0.00\%$	$99.40\pm 0.00\%$
Joint angle error (deg) $\downarrow$	$6.67\pm 0.00$	$6.63\pm 0.01$	$6.19\pm 0.01$	$6.09\pm 0.02$
Joint velocity error (deg/s) $\downarrow$	$27.15\pm 0.03$	$26.92\pm 0.03$	$18.20\pm 0.01$	$17.92\pm 0.03$
Root position error (cm) $\downarrow$	$6.66\pm 0.02$	$7.11\pm 0.15$	-	-

Table 2: Validation metrics comparing MyoFullBody and MyoBimanualArm environments using GMR-Fit retargeting with

N=3

. MyoFullBody uses KINESIS training (972 motions) and testing (108 motions). MyoBimanualArm uses Bimanual training (1770 motions) and testing (312 motions). Root position and root yaw errors are not applicable for MyoBimanualArm environment. See Sec. 5.2 for explanation of dataset selection and Appendix E for the exact definition of each metric element.

Metric	MyoFullBody
Metric	Mocap-Body	GMR-Fit
Success rate (%) $\uparrow$	$79.94\pm 3.85$	$77.67\pm 5.13$
Frame Coverage (%) $\uparrow$	$90.15\pm 2.37$	$90.67\pm 2.92$
Joint angle error (deg) $\downarrow$	$10.67\pm 0.11$	$7.97\pm 0.05$
Joint velocity error (deg/s) $\downarrow$	$40.95\pm 0.18$	$36.51\pm 0.28$
Mean episode length $\uparrow$	$508.68\pm 13.39$	$512.00\pm 16.52$

Table 3: Validation metrics comparing MyoFullBody environments using MoCap-Body vs GMR-Fit retargeting method, with

N=3

using KINESIS testing (108 motions) at 2 billion timesteps. Using Mocap-Body retargeting increases significantly the joint angle and joint velocity error.

Panel of Figure 7 — Figure 7: Representative left lower-limb joint kinematics (hip, knee, ankle, and foot) over a complete gait cycle, comparing experimental human data and MyoFullBody-generated motion. Simulated results were evaluated on five AMASS sequences (KIT/317/walking_medium01–05_poses for walking and KIT/317/walking_run01–05_poses for running), aligned by GRF onset and truncated to a single gait cycle. Only kinematic data are available for [ 57 ] .

Figure 8: Physiological plausibility of synthetic muscle activations during walking. Comparison between gait-cycle-averaged synthetic muscle activations generated by the policy and experimental EMG recordings during level walking. Left: Average muscle–EMG correlation across all recorded muscles and corresponding lower-limb MPJAE. Each triangle represents a human-to-human pair; each dot represents a model-to-human pair. Right: Average activation patterns and correlation values for single muscles.

Figure 9: The 17 mimic sites used for full-body motion imitation, shown with their corresponding anatomical keypoints. Blue balls indicate mimic site locations on the body segments. For MyoBimanualArm, a subset of 6 upper-limb sites plus 1 thorax site is used, with the thorax reference site for computing relative positions.

Figure 10: Overview of the motion retargeting pipeline with either Mimic or GMR. Human shape is first fitted using SMPL and globally scaled to the target model. Motion is then retargeted via inverse kinematics (MuJoCo-based with Mimic or Mink-based with GMR and equality constraints) and interpolated to the control frequency. Post-processing filters mimic sites and corrects artifacts such as ground penetration and floating.

KINESIS (972 motions)
Method	$P_{\text{joint}}$ (%)	$P_{\text{pen}}$ (%)	$D^{\text{max}}_{\text{pen}}$ (m)	$H^{\text{max}}_{\text{float}}$ (m)	$\Delta L^{\text{max}}_{\text{tj}}$ (–)	$P_{\text{tj}}$ (%)	RMSE (m)	$T_{\text{frame}}$ (s)
Mocap-Body retargeting	12.26	0.55	0.002	0.023	0.04	30.14	0.039	0.076
GMR-Fit retargeting	0.27	0.24	0.001	0.021	0.01	3.20	0.025	0.251
Transition (110 motions)
Mocap-Body retargeting	5.67	1.83	0.004	0.012	0.08	62.73	0.030	0.056
GMR-Fit retargeting	0.40	1.64	0.004	0.017	0.07	60.00	0.027	0.142

Table 4: Performance comparison between Mocap-Body retargeting and GMR-Fit retargeting over 2021 motions, including walking, turning, running, jumping, kicking, crawling, fighting, etc. We report eight distinct metrics averaged over all motions: (1) Joint limit violation rate:

P_{\text{joint}}

, averaged across all motions (2) Ground penetration rate:

P_{\text{pen}}

(3) Maximum ground penetration:

D^{\text{max}}_{\text{pen}}

(4) Maximum floating height:

H^{\text{max}}_{\text{float}}

(5) Maximum tendon jump:

\Delta L^{\text{max}}_{\text{tj}}

(6) Tendon jump rate:

P_{\text{tj}}

, out of all motions (7) Root Mean Square Error: RMSE (8) Retargeting speed:

T_{\text{frame}}

. For all metrics, lower indicates better retargeting performance.

Figure 11: Policy observation structure. The state is decomposed into proprioceptive signals (root height and velocity, joint positions and velocities), tendon states, touch info, mimic site relative positions, and motion phase. A history of stacked states is concatenated with the current goal and future goals at regular look ahead intervals. Each goal is defined by root position and velocity deltas and target mimic site relative positions. The previous action $a_{t-1}$ is appended to introduce a first-order autoregressive dependency.

Figure 12: Gated residual policy architecture. Observations are first normalized by running statistics and then passed through

N

residual blocks. Each block applies two Dense–LayerNorm layers with a learnable gate (

w_{l}=\sigma(g_{l})

) on the residual branch and a projection shortcut when dimensions change. If the number of hidden layers is odd, the final layer is processed as a standalone Dense–LayerNorm–SiLU layer.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 13 — Figure 13: Impact of termination threshold on learning performance. A threshold larger than 1 initially shows lower returns and shorter episodes but ultimately achieves superior convergence compared to more restrictive thresholds of 0.25 and 0.5.

Methods	Success Rate ( $\uparrow$ )	J. Pos Err ( $\downarrow$ )	J. Vel Err ( $\downarrow$ )	Root XYZ ( $\downarrow$ )	Root Yaw ( $\downarrow$ )	RPos Err ( $\downarrow$ )	Site Abs ( $\downarrow$ )	Coverage ( $\uparrow$ )	Ep. Len ( $\uparrow$ )	Ep. Ret ( $\uparrow$ )
Baseline	80.76 $\pm$ 5.13	7.94 $\pm$ 0.42	35.97 $\pm$ 1.82	8.47 $\pm$ 0.36	2.63 $\pm$ 0.24	2.85 $\pm$ 0.28	15.37 $\pm$ 0.70	91.10 $\pm$ 2.47	513.4 $\pm$ 14.0	516.3 $\pm$ 26.3
KINESIS (-)	36.63 $\pm$ 2.93	7.87 $\pm$ 0.27	30.09 $\pm$ 0.34	6.91 $\pm$ 0.28	3.53 $\pm$ 0.46	3.15 $\pm$ 0.14	12.90 $\pm$ 0.50	63.65 $\pm$ 0.86	358.4 $\pm$ 4.8	375.1 $\pm$ 8.2
KINESIS (+)	84.77 $\pm$ 3.74	8.29 $\pm$ 0.08	38.02 $\pm$ 0.61	8.13 $\pm$ 0.37	2.82 $\pm$ 0.10	2.93 $\pm$ 0.03	14.84 $\pm$ 0.65	92.05 $\pm$ 1.57	518.8 $\pm$ 8.9	507.7 $\pm$ 7.9
Num Steps = 8	53.09 $\pm$ 2.33	9.64 $\pm$ 0.01	38.77 $\pm$ 0.16	8.13 $\pm$ 0.10	4.64 $\pm$ 0.11	4.43 $\pm$ 0.07	15.54 $\pm$ 0.23	79.07 $\pm$ 1.22	445.5 $\pm$ 6.9	403.2 $\pm$ 4.9
Num Steps = 50	78.70 $\pm$ 3.21	9.76 $\pm$ 0.05	42.36 $\pm$ 0.11	8.32 $\pm$ 0.07	3.09 $\pm$ 0.10	3.18 $\pm$ 0.03	15.23 $\pm$ 0.11	89.94 $\pm$ 1.02	506.8 $\pm$ 5.8	472.3 $\pm$ 6.3
Low Std = 0.2	75.93 $\pm$ 3.34	8.67 $\pm$ 0.03	33.67 $\pm$ 0.18	8.77 $\pm$ 0.19	3.76 $\pm$ 0.11	3.51 $\pm$ 0.03	16.28 $\pm$ 0.36	90.25 $\pm$ 1.21	508.6 $\pm$ 6.8	518.2 $\pm$ 7.4

Appendix

Table 5: Ablation Study on CPU. Evaluation of various hyperparameters on the final performance of the MuscleMimic pipeline. All distance metrics are reported in cm, and angles in degrees. The overall top three methods are highlighted, and the best individual metric across all configurations is bolded.

Methods	Success Rate ( $\uparrow$ )	J. Pos Err ( $\downarrow$ )	J. Vel Err ( $\downarrow$ )	Root XYZ ( $\downarrow$ )	Root Yaw ( $\downarrow$ )	RPos Err ( $\downarrow$ )	Site Abs ( $\downarrow$ )	Coverage ( $\uparrow$ )	Ep. Len ( $\uparrow$ )	Ep. Ret ( $\uparrow$ )
Baseline	82.20 $\pm$ 8.02	7.93 $\pm$ 0.42	36.15 $\pm$ 1.93	8.14 $\pm$ 0.42	2.66 $\pm$ 0.25	2.84 $\pm$ 0.28	14.79 $\pm$ 0.81	91.06 $\pm$ 3.65	514.2 $\pm$ 20.6	517.2 $\pm$ 34.5
KINESIS (-)	37.55 $\pm$ 2.18	7.84 $\pm$ 0.25	30.18 $\pm$ 0.21	6.61 $\pm$ 0.32	3.49 $\pm$ 0.44	3.13 $\pm$ 0.12	12.40 $\pm$ 0.62	63.38 $\pm$ 0.64	357.9 $\pm$ 3.6	374.9 $\pm$ 6.4
KINESIS (+)	85.70 $\pm$ 2.70	8.27 $\pm$ 0.08	38.08 $\pm$ 0.58	7.97 $\pm$ 0.49	2.80 $\pm$ 0.07	2.94 $\pm$ 0.05	14.57 $\pm$ 0.82	92.96 $\pm$ 1.57	524.9 $\pm$ 8.9	513.5 $\pm$ 9.0
Num Steps = 8	52.16 $\pm$ 1.41	9.67 $\pm$ 0.05	38.83 $\pm$ 0.22	8.25 $\pm$ 0.34	4.61 $\pm$ 0.13	4.44 $\pm$ 0.06	15.81 $\pm$ 0.58	77.75 $\pm$ 1.02	439.0 $\pm$ 5.8	397.4 $\pm$ 3.5
Num Steps = 50	79.94 $\pm$ 5.66	9.73 $\pm$ 0.07	42.55 $\pm$ 0.05	8.22 $\pm$ 0.18	3.04 $\pm$ 0.15	3.19 $\pm$ 0.03	15.09 $\pm$ 0.29	91.45 $\pm$ 2.68	516.4 $\pm$ 15.1	481.0 $\pm$ 13.9
Low Std = 0.2	77.16 $\pm$ 0.53	8.64 $\pm$ 0.04	33.58 $\pm$ 0.12	8.17 $\pm$ 0.18	3.74 $\pm$ 0.16	3.49 $\pm$ 0.03	15.24 $\pm$ 0.24	90.26 $\pm$ 0.99	509.7 $\pm$ 5.6	520.1 $\pm$ 5.4

Appendix

Table 6: Ablation Study on GPU. Evaluation of various hyperparameters on the final performance of MuscleMimic pipeline. All distance metrics are reported in cm, and angles in degrees. The overall top three methods are highlighted, and the best individual metric across all configurations is bolded.

Panel of Figure 14 — Figure 14: Comprehensive ablation study results for the MuscleMimic framework (Part I).

Panel of Figure 15 — Figure 15: Comprehensive ablation study results for the MuscleMimic framework (Part II).

Panel of Figure 16 — Figure 16: Example validation of symmetry between left and right muscle-tendon groups of MyoFullBody.

Panel of Figure 17 — Figure 17: Comparison of right-arm muscle moment arm and force–length profiles between the MyoArm model (pre-tuning) and the Mimic-based MyoBimanualArm model (post-tuning).

Panel of Figure 18 — Figure 18: Example validation comparing MyoFullBody muscle moment arms against experimental data from prior studies for selected shoulder, elbow, and lower-limb muscles. Despite inter-individual variability in moment arms and attachment sites, our model’s profiles remain within the reported experimental ranges. (first part)

Panel of Figure 19 — Figure 19: Example validation comparing MyoFullBody muscle moment arms against experimental data from prior studies for selected shoulder, elbow, and lower-limb muscles. Despite inter-individual variability in moment arms and attachment sites, our model’s profiles remain within the reported experimental ranges. (last part)

Body Segment	MyoFullBody (kg)	MyoFullBody (%)	Winter et al., 2009 [ 105 ] (%)
Head (& Neck)	2.4	2.8	8.1
Thorax	18.6	22.1	21.6
Abdomen / Lumbar	9.2	10.9	13.9
Pelvis	18.4	21.8	14.2
Arms (Both)	9.3	11.0	10.0
Legs (Both)	26.9	31.9	32.2

Appendix

Table 7: The MyoFullBody model is compared against biomechanical reference data from [ 105 ] . In the reference studies [ 68 ] , the abdominal segment is defined as spanning T12–L1 to L4–L5, while the thoracic segment spans C7–T1 to T12–L1, and the pelvic segment spans L4–L5 to the greater trochanter. In contrast, the MyoFullBody model defines the abdomen (or hereby lumbar) as L1–L5, with the thorax defined as C7 until T12 - L1. In addition, the MyoFullBody does not represent the neck as a separate segment, and its mass is therefore included within the head segment. The sacrum is classified as part of the pelvis body segment in the table below. These differences in segment definitions should be considered when interpreting discrepancies in segmental mass percentages.

Category	Parameter	Value
Environment	Number of parallel environments	8,192
	Episode horizon	1,000 steps
	Total training timesteps	15.87 billion
	Backend	MuJoCo Warp
Network Architecture	Actor hidden layers	[2048, 4096, 4096, 4096, 4096, 4096, 2048, 1024, 512]
	Critic hidden layers	[2048, 4096, 4096, 4096, 4096, 4096, 2048, 1024, 512]

Appendix

Table 8: Training hyperparameters for MyoFullBody motion imitation pretrained checkpoint.

Category	Parameter	Value
Environment	Number of parallel environments	8,192
	Episode horizon	1,000 steps
	Total training timesteps	2.048 billion
	Backend	MuJoCo Warp
Network Architecture	Actor hidden layers	$16$ layers: $12{\times}1024\to 3{\times}2048\to 1024$
	Critic hidden layers	$16$ layers: $12{\times}1024\to 3{\times}2048\to 1024$

Appendix

Table 9: Training hyperparameters for MyoBimanualArm motion imitation pretrained checkpoint.

Explore similar work

Mar 31, 2026

Scaling Whole-Body Human Musculoskeletal Behavior Emulation for Specificity and Diversity

The embodied learning of human motor control requires whole-body neuro-actuated musculoskeletal dynamics, while the internal muscle-driven processes underlying movement remain inaccessible to direct measurement. Computational modeling offers an alternative, but inverse dynamics methods struggled to resolve redundant control from observed kinematics in the high-dimensional, over-actuated system. Forward imitation approaches based on deep reinforcement learning exhibited inadequate tracking performance due to the curse of dimensionality in both control and reward design. Here we introduce a large-scale parallel musculoskeletal computation framework for biomechanically grounded whole-body motion reproduction. By integrating large-scale parallel GPU simulation with adversarial reward aggregation and value-guided flow exploration, the MS-Emulator framework overcomes key optimization bottlenecks in high-dimensional reinforcement learning for musculoskeletal control, which accurately reproduces a broad repertoire of motions in a whole-body human musculoskeletal system actuated by approximately 700 muscles. It achieved high joint angle accuracy and body position alignment for highly dynamic tasks such as dance, cartwheel, and backflip. The framework was also used to explore the musculoskeletal control solution space, identifying distinct musculoskeletal control policies that converge to nearly identical external kinematic and mechanical measurements. This work establishes a tractable computational route to analyzing the specificity and diversity underlying human embodied control of movement. Project page: https://lnsgroup.cc/research/MS-Emulator.

Yunyue Wei, Chenhui Zuo, Shanning Zhuang +3

Tsinghua University.

Whole-Body Control Human Motion Generation RL Control Imitation Learning

Ask PDF

Sep 18, 2026

Beyond Kinematics: Benchmarking Simulation Fidelity for Muscle-Driven Imitation Learning

In this work, we conduct a systematic comparison of two state-of-the-art motion-imitation reinforcement learning (MIRL) pipelines, one built on SCONE/HyFyDy and one built on MuJoCo/MyoSim. HyFyDy emphasizes physiological realism through detailed musculotendon modeling, while MuJoCo prioritizes computational efficiency and scalable policy learning. While recent work has demonstrated that both pipelines reproduce human kinematics with high fidelity, it remains unclear if they accurately capture the underlying neuromuscular behavior that produced the movement. This limitation is particularly important for robotic assistive-device design and control, where outcome measures such as muscle activation patterns and metabolic cost are often used as optimization targets. To conduct a systematic comparison, our work compares both pipelines using a common set of human motion-capture and electromyography (EMG) measurements. The results find that while both pipelines produce similar kinematics with relative accuracy, the muscle activations from HyFyDy are more aligned with the experimental EMG, as supported by the average pooled (RMSE, r) values for muscle activations from HyFyDy and MuJoCo: (0.164, 0.4) and (0.344, 0.11), respectively. While we conclude that the more advanced physiological realism of HyFyDy currently makes it more suitable for musculoskeletal modeling, both require further development to bring physiological realism to GPU-parallelizable simulation environments and advance robotic assistive device design.

Ayah G. Ahmad, Claire E. Borden, Maegan Tucker

Dynamic Mobility Lab at Georgia Tech, Atlanta, U.S.

Robotics Simulation Imitation Learning

Ask PDF

Jun 3, 2026

M3imic: Learning a Versatile Whole-Body Controller for Multimodal Motion Mimicking

Building a general-purpose whole-body controller is essential for enabling diverse motion capabilities in humanoid robots across a wide range of downstream tasks, including locomotion and loco-manipulation. Different tasks rely on distinct motion reference modalities: locomotion primarily depends on coordinated robot joint trajectories, whereas manipulation requires precise end-effector trajectory tracking. Existing methods often overlook the representational mismatch between dense robot joint angles and sparse end-effector poses. To address this, we propose Multi-Modal Mimic (M3imic), a versatile multi-modal whole-body control framework that unifies heterogeneous motion reference modalities, including robot joint angles, human pose trajectories, and end-effector poses, using modality-specific encoders to map them into a shared latent space. Leveraging large-scale reinforcement learning in the simulator, we train a single policy that achieves sim-to-real transfer across multiple motion reference modalities without modality-specific retraining. Extensive simulation and real-world experiments on the Unitree G1 robot are conducted to evaluate the proposed framework. In simulation, the policy achieves a peak success rate of 98.42% on an unseen test dataset, demonstrating its exceptional generalization capability. The code is available at https://github.com/Renforce-Dynamics/MultiModalWBC

Zuxing Lu, Ziang Zheng, Yao Lyu +7

School of Automation, Southeast University, Nanjing 210096, China · School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China · Department of Robotics, Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi 7909, UAE

RL for Robotics Humanoid Robot Control Whole-Body Control Robot Imitation Learning Multimodal Learning

Ask PDF