GRASP: Gradient-Guided Sampling for Contact-Rich Physics-Based Motion Retargeting
AuthorsYang Liu, Tong Wang, Yiming Bao, Heng Yang
Affiliations[
GRASP uses past simulation rollouts to guide robot motion retargeting, producing contact-rich movements that work in simulation and on real hardware.
Key results
Success rate across 1352 bimanual demonstrations.
Object-tracking score across 5400 trajectories.
Mean reward with GRASP-refined references, versus 21.7 for kinematic references.
Seconds to reach cost 2504; gradient-based TSMC took 2069 seconds to reach cost 2534.
What the paper found
GRASP combines population-based exploration with gradient-guided refinement for physics-based motion retargeting, without differentiating through the simulator. It archives evaluated control sequences and costs, including rejected proposals, then fits local quadratic models in a reduced temporal basis; their recovered gradients guide Metropolis-adjusted Langevin proposals, while Metropolis–Hastings correction preserves the original target distribution. GRASP-MPC extends this optimizer to long demonstrations with receding-horizon replanning and warm-started plans. Using MuJoCo and NVIDIA A100 GPUs, the method finds contact-rich motions with a shared tracking objective and no task-specific contact rewards. On 1352 bimanual OakInk2 demonstrations, it achieves 92.8% success; on 5400 HOT3D trajectories, it reaches 81.54% ADD-AUC. In pen-spinning policy learning, PPO trained on refined references raises mean episode reward from 21.7 to 45.9. On the acrobot benchmark, GRASP-TO reaches a cost of 2504 in 46 seconds, compared with 2534 in 2069 seconds for gradient-based TSMC. The paper also demonstrates dexterous manipulation on physical hands and closed-loop humanoid behaviors, including crawling and box handling on a Unitree G1.
Original abstract
We study physics-based retargeting of human demonstrations for robot learning: finding robot controls that reproduce reference motions under contact dynamics. We propose GRASP (Gradient Recovery from Archived Samples for Proposals), which combines sampling-based exploration with gradient-guided refinement without requiring a differentiable simulator. GRASP-TO extends tempered sequential Monte Carlo by fitting local quadratic cost models to archived rollouts and using their gradients to guide Markov chain Monte Carlo rejuvenation. Metropolis-Hastings correction preserves the target distribution for each fixed proposal kernel, and a population analysis connects the recovered gradients to Gaussian smoothing. GRASP-MPC scales this optimizer to long demonstrations through warm-started receding-horizon planning. Experiments on dexterous hands and humanoids show accurate retargeting with a shared tracking objective and no task-specific contact rewards, including 92.8% success on 1352 bimanual OakInk2 demonstrations. The refined references improve PPO tracking in pen spinning and humanoid-object interaction. We also demonstrate dexterous manipulation and whole-body humanoid motion on real hardware.
Read the original paperMore in Robotics
Browse all 68 papers →ELA: End-to-End Latent Action Policy Learning from Video Demonstrations
Cheng-Chun Hsu, Peter Stone
ELA turns plentiful robot videos into useful control experience by jointly learning latent actions, policies, and action decoders.
PERA: Plug-in Experts for Robotic Actions in Continual Whole-Body Control
Shangchao Hu, Yadan Sun, Xibo Yu, Wei Huang
PERA teaches humanoid robots new skills by plugging small experts into an existing controller without disrupting its original motions.
Alloy: A Multi-Material Physics Engine for Robot Learning
Pranay Gupta, Deepanshu Rohilla
Alloy brings robots, cloth, cables, liquids, and other materials into one GPU-capable simulator for robot learning.