Robotics Research Hub
Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking

Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking

Pranit Mohnot, Christian Helten, Daniele Gammelli, Marco Pavone · N/A · 2026

reinforcement-learning control
PDF arXiv

Framework

N/A

License

N/A

Stars

N/A

Summary

Safe whole-body motion is essential for deploying humanoid robots in unstructured environments.

Abstract Summary

Safe whole-body motion is essential for deploying humanoid robots in unstructured environments. Modern humanoid control commonly separates reference specification from execution, with a planner, teleoperator, or motion generator providing a reference that a reinforcement-learning policy tracks through dynamically feasible whole-body control. Runtime safety filters, such as control barrier functions (CBFs), offer a promising approach for enforcing newly introduced constraints via interventions on the tracker's outputs. We show, however, that treating the tracking policy and safety filter independently induces fundamental mismatches, as filtering alters both the executed actions and the induced state distribution. We study this policy-filter interface through case studies that isolate dynamics, objective, and information mismatches, highlight their root causes, and use these insights to develop CoFiT (Constrained Filter-aware Tuning), a filter-aware fine-tuning method for pretrained trackers. Across diverse constraint scenes, CoFiT reduces violation time relative to filter-only training by 91% on TWIST2 and 21% on SONIC, while requiring smaller safety filter corrections. On Unitree G1 hardware, CoFiT reduces violation time by 83% for TWIST2 and completes every trial without operator intervention, whereas 50% of baseline trials require an operator stop. Together, these results provide actionable insights into policy-filter interactions and establish design principles for integrating learned trackers with runtime safety filters.

Key Points

  • Safe whole-body motion is essential for deploying humanoid robots in unstructured environments
  • Modern humanoid control commonly separates reference specification from execution, with a planner, teleoperator, or motion generator providing a reference that a reinforcement-learning policy tracks through dynamically feasible whole-body control.
  • Runtime safety filters, such as control barrier functions (CBFs), offer a promising approach for enforcing newly introduced constraints via interventions on the tracker's outputs.
  • We show, however, that treating the tracking policy and safety filter independently induces fundamental mismatches, as filtering alters both the executed actions and the induced state distribution.
  • We study this policy-filter interface through case studies that isolate dynamics, objective, and information mismatches, highlight their root causes, and use these insights to develop CoFiT (Constrained Filter-aware Tuning), a filter-aware fine-tuning method for pretrained trackers.

Preliminaries

In this section, we first briefly review the relative-degree-two CBF formulation used to encode safety constraints (Section III-A ), and then describe the whole-body filtering framework through which these constraints modify the tracking policy’s nominal commands (Section III-B ). A control barrier function expresses a state constraint as a condition on the control input [ 5 ] .

Preliminaries — Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking
Fig. 2: Tracking with the filter on (top) and off (bottom) with the same base policy and scene. White: robot. Red: reference. Cyan: barrier violation.

Failure Modes of Independently Designed Policies and Safety Filters

Base trackers were trained with a nominal unconstrained model of the actuators, but applying filters at runtime introduces a distribution shift that impacts the tracking and stability performance of the underlying policy. We distinguish three interacting mismatches in the composed policy-filter loop, regarding dynamics, objective, and information.

CBFs produce endogenous and autocorrelated perturbations on unfiltered dynamics.

Failure Modes of Independently Designed Policies and Safety Filters — Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking
Fig. 3: CoFiT observation conditioning

Filter-Aware Policy Design

CoFiT (illustrated in Fig. 3 ) fine-tunes whole-body trackers to cooperate with a safety filter while maintaining performance. We retain the base tracker’s MLP structure and native tracking rewards as a backbone, with the ConstrainedMimic whole-body filter active throughout training and deployment [ 7 ] . As motivated in Sec.

IV-C , we endow the base policy with information regarding which directions would lead to violation, and how urgent the constraint is.

Filter-Aware Policy Design — Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking
Figure from the paper. Source: Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking.

Experiments

We evaluate CoFiT through four outcomes: realized safety, cooperation with the filter, motion quality, and reference tracking. We first isolate design choices in a simplified setting, then validate on more realistic scenes, a second tracker, and hardware. We provide the tracker with fixed world-frame trajectories that cannot replan in response to tracking errors or safety interventions.

Our scenes are therefore designed as near-misses or mild interferences requiring local readjustment rather than rerouting.

Experiments — Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking
Fig. 5: Generated constraint scenes with overlaid reference poses and task-relevant obstacles highlighted in blue.

Sources and Demonstrations

Figures are reproduced from Mohnot et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.

Share

Related Papers