Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation

summary

The gist

The gist: Compliant GMT (C-GMT) is a modular framework that unifies general motion tracking with adjustable whole-body compliance through residual reference adaptation, achieving both competitive

In short

Compliant GMT creates a system that allows humanoid robots to perform any general motion while safely responding flexibly to external forces. It combines a robust general motion tracker with a residual adapter that modifies motions based on proprioception and desired compliance. This approach achieves both accurate tracking and safe, compliant whole-body responses.

Key concepts

Force-Robust General Motion Tracker: This component faithfully executes reference motions even when subjected to external loads. It is trained to be resilient to forces by finetuning its decoder using simulated payloads and sustained external forces within a specific force budget. This ensures accurate motion tracking under real-world disturbances.
Compliant Teacher: A privileged teacher generates ideal compliant reference motions based on an applied external force and commanded stiffness. It uses a spring law to determine the desired position of a forced link and an inverse kinematics (IK) solver to create a whole-body reference prioritizing balance and foot placement.
Residual Adapter: This learned module modifies the motion token generated by the general tracker to introduce compliance. It is conditioned on a history of proprioceptive observations and previous actions, learning how to adjust the reference based on actual forces encountered during deployment.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 20 episodes
Retrieval-Augmented Generation (RAG) — explained in 19 episodes
Federated Learning — explained in 17 episodes
Flow Matching — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes

This episode discusses

The paper

Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation · Read on arXiv

Venki Parsuram, Michelle Wang, Jin-Hyuk Hong, Pulkit Agrawal

We propose Compliant GMT (C-GMT), a modular framework for full-body compliant responses in general motion tracking. As humanoid robots move into real world deployment, they not only need to perform any motion but also respond safely and robustly to interaction forces. Existing general motion tracking policies can imitate diverse motions but are trained to be stiff, leading to unexpected and unsafe behaviors under disturbance. Prior work in compliant humanoid control trades off compliance and generality. There exist examples of full-body compliant responses, but these are heavily curated to single-reference policies. On the other hand, general motion trackers (GMT) are trained to rigidly imitate massive motion capture datasets. Our key insight is to leverage the generality of pretrained GMTs for motion tracking while introducing compliance from a residual adapter that conditions on a history of proprioception. This enables our approach, C-GMT, to achieve both general motion tracking and whole-body compliant responses. Compared to state-of-the-art general motion trackers and single-motion compliant policies, C-GMT demonstrates competitive performance in both motion tracking accuracy and compliance adherence. We further demonstrate improved safety, robustness, and error tolerance in simulation and real-world G1 experiments.

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: Today's paper: "Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation".

Dev: The gist: Compliant GMT (C-GMT) is a modular framework that unifies general motion tracking with adjustable whole-body compliance through residual reference adaptation,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, wrapping up Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation, they’ve built a modular system that unifies general motion tracking with adjustable whole-body compliance using this residual reference adaptation method. They propose a proprioceptive residual adapter distilled from a force-privileged teacher to generate compliant references for the general motion tracker.

Dev: The paper really highlights that this framework allows compliance to be specified independently of the core motion tracking objective, which is what separates it from previous work that had to trade off compliance for generality.

Taro: What does this mean in practice? It means we can take a general motion tracker and make it safe and robust under interaction forces without completely losing its ability to perform diverse, dynamic motions.

Rosa: Right. It suggests that low-level compliance built into the motion-tracking layer actually improves the robustness of the higher-level policies without sacrificing how well they track the original task.

Dev: The authors show that C-GMT maintains tracking performance close to SONIC, with only a modest increase in global position error, and they also show improved safety and error tolerance in simulation and real-world G1 experiments.

Taro: So, the main implication for me is that when you're building these systems for real world deployment, you don't have to choose between being flexible enough to do any motion or being safe enough not to break when things go wrong.

Rosa: That’s the point. The title itself is quite descriptive: Compliant GMT: Making General Motion Tracking Compliant via Residual Reference Adaptation. It tells us exactly what they achieved by showing how you can adapt the reference motion using proprioception and stiffness commands in a way that's effective during deployment.

Conclusion: Rosa: So, C-GMT is this modular framework they proposed to make general motion tracking compliant by using residual reference adaptation.

Dev: Yeah, it’s about decoupling the motion tracking from the compliance part so you can tune them separately.

Taro: So instead of one system trying to do both things at once, you have a force-robust tracker and then this residual adapter that modifies the reference based on what the robot is actually feeling.

Rosa: Right. The authors build this teacher privileged system to generate those compliant references first so they can verify the feasibility before training the student part.

Dev: And for deployment, that teacher behavior gets distilled into a residual adapter conditioned on proprioception and a commanded stiffness value.

Taro: It sounds like they’re taking the general motion tracking ability of a pre-trained system and layering on this real-time compliance layer that reacts to the robot's own feedback.

Rosa: Exactly. And their results show it does both things well, matching the accuracy of state-of-the-art general trackers while also showing better safety during dynamic motions in experiments.

Dev: The paper suggests that you can improve robustness to sustained external forces by allowing the whole body to redistribute the disturbance instead of just resisting it rigidly.

Taro: That means if you push against something hard, the robot doesn't just get stiff and break or jerk; it yields in a controlled way across its structure.

Rosa: It shifts thinking away from single-motion compliant policies that are only good for one specific movement, toward something that works for any motion while being safe when things go wrong.

Dev: This implies that low-level compliance built into the core tracking mechanism can actually help the whole system be more robust at a higher level.

Taro: It’s about making sure those high-level policies aren't just brittle imitations of data but have some inherent physical safety built in.

Rosa: So, C-GMT is this new way to get general motion tracking that doesn't sacrifice the ability to react safely when interacting with the real world.

Dev: And they show it works on hardware like the Unitree G1 in simulation and in real-world G1 tests, which is important for seeing how it handles those unpredictable physical situations.

Taro: It gives us a way to test general motion tracking under load and see how gracefully the system recovers when things go wrong.

More episodes

← Home