SoftMimic: Learning Compliant Whole-body Control from Examples

TL;DR

SoftMimic tackles the challenge of safe, compliant whole-body control for humanoids by learning from augmented compliant trajectories generated offline via differential inverse kinematics. A PPO-based policy learns to map proprioceptive state and the original motion reference to the corresponding augmented targets, enabling controllable compliance with external forces while preserving tracking fidelity. The Compliant Motion Augmentation pipeline ensures feasible, diverse demonstrations and, combined with reward shaping, yields robust generalization to unseen tasks and disturbances, demonstrated in simulation and on a real Unitree G1. This approach advances practical humanoid interaction by balancing stiffness and safety, and it opens avenues for dynamic stiffness adaptation and richer data-driven compliance in future work.

Abstract

We introduce SoftMimic, a framework for learning compliant whole-body control policies for humanoid robots from example motions. Imitating human motions with reinforcement learning allows humanoids to quickly learn new skills, but existing methods incentivize stiff control that aggressively corrects deviations from a reference motion, leading to brittle and unsafe behavior when the robot encounters unexpected contacts. In contrast, SoftMimic enables robots to respond compliantly to external forces while maintaining balance and posture. Our approach leverages an inverse kinematics solver to generate an augmented dataset of feasible compliant motions, which we use to train a reinforcement learning policy. By rewarding the policy for matching compliant responses rather than rigidly tracking the reference motion, SoftMimic learns to absorb disturbances and generalize to varied tasks from a single motion clip. We validate our method through simulations and real-world experiments, demonstrating safe and effective interaction with the environment.

Figures (8)

• Figure 1: Soft Whole-body Control via Compliant Motion Augmentation.Left: Given an original reference motion ($q_\text{ref}$) and a specified interaction (external wrench $W_\text{ext}$ and stiffness $K_\text{robot}$), our offline data generation stage uses an IK solver to generate a kinematically feasible and stylistically consistent compliant motion ($q_\text{aug}$). Right: During online training, a policy learns to reproduce this behavior. It observes the robot's proprioceptive state and the original reference ($q_\text{ref}$), but is rewarded for matching the augmented target ($q_\text{aug}$). This forces the policy to implicitly infer the external wrench and react appropriately, resulting in a robot that can controllably comply with generalized unanticipated perturbations. In the graphic, we only annotate translational forces and displacements for ease of interpretation, but the analogous rotational quantities are also simulated.
• Figure 2: Stiffness adherence. The humanoid’s effective translational stiffness tracks the commanded stiffness over a wide range. We apply external forces to the hands of a standing robot in simulation and report the median force–displacement ratio for a single SoftMimic policy across stiffness commands (log–log scale). A stiff motion-tracking baseline maintains an approximately constant effective stiffness of 550Nm under identical conditions. Data augmentation further improves adherence, especially at low stiffness (Fig. \ref{['fig:aug_benefit']}).
• Figure 3: SoftMimic reduces collision forces across various motions in unseen environments. The bar chart compares the maximum contact force generated by our policy (at low and high stiffness) and the stiff baseline across three challenging scenarios involving unexpected contact. In all cases, the compliant policy operating at a low stiffness significantly reduces interaction forces, enhancing safety.
• Figure 4: Stiffness modulation controls collision forces. The plot shows the contact force over time as the robot's hand collides with a tower of blocks. By commanding different stiffness levels, our policy can produce low, controlled forces (blue) or high, potentially destructive forces (red), showcasing a direct trade-off between safety and posture tracking accuracy.
• Figure 5: SoftMimic enables generalization to unseen objects and disturbance scenarios. Using a single motion reference designed for a 20cm box, our policy can successfully pick up boxes of increasing width. By commanding the same low stiffness, the robot not only successfully picks up different sized boxes with a consistent, gentle squeezing force, but it also is able to safely handle collisions with misaligned boxes. In contrast, the stiff baseline generates large and uncontrolled force spikes, risking damage to the object or robot.