ContactMimic: Humanoid Object Interaction via Contact Control

Noosaga Placements

• Robot Learningsubfield95%
  The paper introduces a learning framework (ContactMimic) that trains a policy via reinforcement learning with contact-following rewards and trajectory augmentation. The core contribution is a learned policy that decouples contact behavior from keypoint geometry.
  We present CONTACTMIMIC, a learning framework that tracks explicit partlevel binary contact commands alongside keypoint trajectories.The resulting policy successfully decouples contact behavior from keypoint geometry, and achieves precise physical contact as well as contact-controllabilityWe train one policy per motion with PPO [44] in Isaac Lab [43] across 4096 parallel environments at 50 Hz.
• Deep Reinforcement Learningframework90%
  The paper uses PPO, a deep reinforcement learning algorithm, to train the contact-conditioned policy. The reward function is designed to encourage contact following, and the policy is trained in simulation.
  We train one policy per motion with PPO [44] in Isaac Lab [43] across 4096 parallel environments at 50 Hz.PPO hyperparameters are listed in Table 5.
• Robot Controlsubfield80%
  The paper focuses on controlling a humanoid robot to make or suppress physical contact based on a contact command. The policy outputs joint angles that are tracked by a PD controller, and the contact-aware rewards directly influence the control behavior.
  The policy outputs target joint angles that are converted into torques using a PD controller.Our proposed contact and keypoint tracker, πθ (at |pt , k̄t , c̄t ), takes as input proprioception pt , reference keypoint position targets k̄t , and a binary reference contact label map c̄tThe policy follows the commanded contact, outperforms a keypoint-only baseline on contact metrics at comparable tracking accuracy
• Learning-Based Robot Controlframework80%
  The paper uses a learning-based approach to robot control. The policy is a neural network that outputs joint angles, which are then tracked by a PD controller. The contact-aware rewards directly shape the control behavior.
  Our proposed contact and keypoint tracker, πθ (at |pt , k̄t , c̄t ), takes as input proprioception pt , reference keypoint position targets k̄t , and a binary reference contact label map c̄tThe policy outputs target joint angles that are converted into torques using a PD controller.
• Robot Manipulationsubfield70%
  The paper evaluates on object interaction tasks such as wiping a board, sitting on a chair, pushing furniture, and picking up a box. These are manipulation tasks that require making contact with objects.
  Keypoint tracking alone is insufficient for object interaction tasks such as sitting on a chair, wiping a board, or pushing furniturecontact conditioning lets us complete object manipulation tasks (e.g. lifting a box), without needing task specific rewards.For free-object motions we also report the object displacement (m), how much the object travels from its initial pose; larger values indicate that the policy successfully manipulates the object.
• Imitation Learningframework70%
  The paper builds on imitation learning from human motion data. It retargets human-object interaction clips to the robot and uses them as reference trajectories for training. The policy learns to imitate both keypoint and contact patterns.
  We leverage human object interaction datasets. Rather than just extracting and retargeting keypoint trajectories, we additionally extract and retarget contact patterns.We use OmniRetarget [29] to retarget a given human-object interaction clip to the G1 humanoid, producing a reference robot configuration trajectory q̄1:T
• Reinforcement Learningsubfield60%
  The policy is trained using PPO, a reinforcement learning algorithm. The reward function includes contact-aware terms that guide the policy to match commanded contact patterns.
  We train one policy per motion with PPO [44] in Isaac Lab [43] across 4096 parallel environments at 50 Hz.The total reward decomposes into standard regularization terms and keypoint-tracking terms inherited from prior work (described in supp.), and contact-aware rewards introduced in this paper
• Learning-Based Manipulationframework60%
  The paper addresses manipulation tasks such as picking up a box and wiping a board. The learning-based approach is applied to manipulation, and the contact conditioning enables task completion without task-specific rewards.
  contact conditioning lets us complete object manipulation tasks (e.g. lifting a box), without needing task specific rewards.For free-object motions we also report the object displacement (m), how much the object travels from its initial pose; larger values indicate that the policy successfully manipulates the object.

Abstract

Keypoint tracking alone is insufficient for object interaction tasks such as sitting on a chair, wiping a board, or pushing furniture, where the robot can reach the correct pose without making meaningful physical contact with the object. We present CONTACTMIMIC, a learning framework that tracks explicit partlevel binary contact commands alongside keypoint trajectories. CONTACTMIMIC is made possible through the use of contact-following rewards and a trajectory augmentation scheme aimed at breaking the correlations between keypoint trajectories and contact labels. The resulting policy successfully decouples contact behavior from keypoint geometry, and achieves precise physical contact as well as contact-controllability (produce or suppress contact during deployment as desired). Simulation experiments across 10 diverse human-object interaction motions confirm that CONTACTMIMIC exhibits contact controllability that enables it to complete manipulation tasks without task-specific rewards, while also outperforming keypoint-only trackers on contact-relevant tasks. Ablations confirm the necessity of the proposed trajectory augmentation scheme and sim2real deployment validates contact controllability in the real world across 5 different motions. Video results are available on https://lixinyao11.github.io/contactmimic-page/.

Paper Context