Paper
ArXiv (Updating In Progress)
Paper
ArXiv (Updating In Progress)
CHIP enables humanoid robots to perform complex, long-horizon tasks that require adaptive compliance transitions — switching between compliant and stiff modes on-the-fly based on task demands.
Compliant but agile box holding -- maintain stable bimanual surface contact while running
Stiff cap open → Hybrid holding+writing → Compliant wiping
Compliant pulling → Stiff door open → Compliant box grasping
Compliant box holding + Leg control -- Step on lid paddle before dropping box into trashcan
Recent progress in humanoid robots has unlocked agile locomotion skills including backflipping, running, and crawling. Yet it remains challenging for a humanoid robot to perform forceful manipulation tasks such as moving boxes, wiping, and pushing a cart. We propose humanoid natural Compliant control through Hindsight Perturbation (CHIP), a method that plugs into any humanoid motion tracking framework, allowing control over robot end-effector stiffness while still agile to track any reference motion. CHIP is easy to use and requires neither data augmentation nor additional reward tuning. We show that a generalist motion tracking controller trained with CHIP can finish a diverse set of forceful manipulation tasks, which require different end-effector compliance, such as multi-robot collaboration, wiping, and door opening.
Different tasks require different levels of compliance. Being too stiff or too compliant can both lead to failure. CHIP enables adaptive compliance control — steering compliance behavior based on task requirements.
Being too stiff causes unstable grasping. Compliance allows the robot to maintain stable contact.
Stable grasping with appropriate compliance
Unstable to maintain contact for two hands -- too stiff
Some tasks require stiff mode to generate sufficient force — compliance behavior alone is not enough.
Compliant → Stiff → Compliant
Cannot lift — too compliant/not strong enough
Strong force generation and accurate following
Not strong enough — fails to re-orient
Suits for this task — successful re-orientation
Forceful flipping of a large box
Heavy objects require a stiffer controller to generate sufficient force for lifting,
whereas a more compliant controller makes it easier to maintain stable contact.
We need to balance stiffness and compliance continuously based on task demands.
Compliance 10/k = 0.3
Stable grasping
Compliance 10/k = 0.05
Forceful lifting (~10 lbs)
Compliance 10/k = 0.4
Stable grasping
This continuous demonstration shows how users can adjust compliance in real-time based on box contents:
Empty box (0.5) → 2 Wipes (0.3) → Wipes + Dumbbell (0.2)
CHIP achieves continuous compliance-aware training through a fundamentally lighter approach: hindsight perturbation. Instead of synthesizing perturbation data via inverse kinematics (as in SoftMimic) or tuning additional reward terms, CHIP simply treats the original reference motion as if it were already the result of a perturbation and reconstructs sparse tracking targets on the fly during RL training. No reference motion modification, no data augmentation, no extra reward tuning — making CHIP significantly more scalable than prior methods.
Because CHIP's compliance command is grounded in a physically meaningful Hooke's law formulation, appropriate levels can be inferred by VLMs like Gemini from just a few user-provided anchor examples (e.g., 0.5 for an empty box, 0.2 for box with about 7 lbs of weight). Gemini can then predict suitable compliance for unseen objects directly from an image — making the tuning process intuitive and nearly effortless.
Stable grasping with Gemini's prediction
Stable grasping
Unstable — too compliant under perturbations
Stable grasping
Unstable grasping due to stiff and competing contact forces
Unable to lift - insufficient force generation
Stable grasping
Stable grasping
When CHIP's compliance spectrum meets local vs. global tracking controllers, we unlock 3 distinct human-robot interaction modes: when compliance is low, the robot stiffly resists perturbation and accurately tracks the global reference; when compliance is high, the robot yields to perturbation while still returning to the global trajectory; and with a damper-only controller on top of CHIP, the robot follows human guidance freely with infinite compliance.
Low compliance — robot resists perturbation and tracks the global reference
High compliance — robot yields to perturbation while still tracking global reference
Damper-only controller — infinite compliance, robot follows human guidance freely
In global tracking setting, CHIP enables compliant multi-robot grasping using SpringGrasp planning. Robots generate pre-grasp trajectories to approach objects collaboratively.
Successful collaborative grasp
Fails to grasp — fail to apply enough force stably
Successful collaborative grasp
Fails to grasp — stiff and competing contact forces
After grasping, robots move the object together following keyboard commands.
Coordinated movement with compliant control
We finetune a Vision-Language-Action (VLA) model GR00T N1.5 on data collected with CHIP for autonomous manipulation tasks with adaptive compliance.
Reactive continuous rollouts with compliant contact (w task success rate: 80%)
A successful run with adaptive compliance (w task success rate: 60%)
A successful run with adaptive compliance (w task success rate: 90%)
First 10 consecutive runs from 20 evaluations (w task success rate: 90%)
@article{chen_cao_2025chip,
title={CHIP: Learning Adaptive Compliance for Humanoid Control through Hindsight Perturbation},
author={Chen, Sirui and Cao, Zi-ang and Luo, Zhengyi and Castañeda, Fernando and Li, Chenran and Wang, Tingwu and Yuan, Ye and Fan, Linxi and Liu, C Karen and Zhu, Yuke},
journal={arXiv preprint arXiv:2512.14689},
year={2025}
}