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

Pranit Mohnot; Christian Helten; Daniele Gammelli; Marco Pavone

cs.ROOct 1, 2026

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

Authors: Pranit Mohnot, Christian Helten, Daniele Gammelli, Marco Pavone

Organizations: Stanford University · Technical University of Munich · Italian Institute of Artificial Intelligence (AI4I) · NVIDIA

Abstract

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.

Figures & tables

Fig. 1: Whole-body tracking with an active safety filter (avoid the ball). The continued-training baseline (left) and CoFiT (right) track the same reaching reference. Five of ten baseline trials required an operator stop; CoFiT completed all ten. Overlaid poses are successive stills from one trial.

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.

Fig. 3: CoFiT observation conditioning

Figure 4

Backbone	Method	Viol.	Max depth	Correction	World MPJPE	TV	Resid. acc.
		(%)	(p95 cm)	(rad RMS)	(cm)	(rad/step)	(rad/s 2 RMS)
TWIST2	CoFiT	1.71 ${}_{\mbox{\scriptsize 0.10}}$	8.27 ${}_{\mbox{\scriptsize 0.55}}$	0.188 ${}_{\mbox{\scriptsize 0.003}}$	4.33 ${}_{\mbox{\scriptsize 0.06}}$	0.174 ${}_{\mbox{\scriptsize 0.003}}$	18.8 ${}_{\mbox{\scriptsize 0.3}}$
	Observations-only	5.01 ${}_{\mbox{\scriptsize 0.22}}$	13.50 ${}_{\mbox{\scriptsize 0.27}}$	0.251 ${}_{\mbox{\scriptsize 0.007}}$	3.54 ${}_{\mbox{\scriptsize 0.04}}$	0.197 ${}_{\mbox{\scriptsize 0.007}}$	18.3 ${}_{\mbox{\scriptsize 0.5}}$
	Penalties-only	15.48 ${}_{\mbox{\scriptsize 0.94}}$	11.34 ${}_{\mbox{\scriptsize 0.64}}$	0.343 ${}_{\mbox{\scriptsize 0.013}}$	5.45 ${}_{\mbox{\scriptsize 0.24}}$	0.180 ${}_{\mbox{\scriptsize 0.014}}$	19.3 ${}_{\mbox{\scriptsize 0.4}}$
	Filter-only	19.31 ${}_{\mbox{\scriptsize 0.58}}$	14.77 ${}_{\mbox{\scriptsize 0.28}}$	0.458 ${}_{\mbox{\scriptsize 0.016}}$	3.77 ${}_{\mbox{\scriptsize 0.08}}$	0.259 ${}_{\mbox{\scriptsize 0.014}}$	19.6 ${}_{\mbox{\scriptsize 0.1}}$
	Unfiltered training	23.75 ${}_{\mbox{\scriptsize 0.27}}$	15.58 ${}_{\mbox{\scriptsize 1.05}}$	0.857 ${}_{\mbox{\scriptsize 0.023}}$	4.08 ${}_{\mbox{\scriptsize 0.03}}$	0.212 ${}_{\mbox{\scriptsize 0.007}}$	24.4 ${}_{\mbox{\scriptsize 0.8}}$

TABLE II: Constrained tracking in generated scenes. Mean with sample SD in subscripts across training seeds.

Fig. 5: Generated constraint scenes with overlaid reference poses and task-relevant obstacles highlighted in blue.

Fig. 6: Hardware examples with the filter active. (a) SONIC overlay comparison. The TWIST2 overlay appears in Fig. 1 . (b,c) RMS keypoint acceleration summarizes acceleration of tracked body points, measured during the pictured trials.

Backbone	Method	(%)	(rad RMS)	(cm)	(rad/step)	(rad/s 2 RMS)	(rad/s RMS)	required †
		Violation	Correction	World MPJPE	Command TV	Residual accel.	Arm vel. error ∗	E-stops
TWIST2	CoFiT	$8.13_{\mbox{\scriptsize 1.98}}$	$0.449_{\mbox{\scriptsize 0.055}}$	$7.55_{\mbox{\scriptsize 1.04}}$	$0.381_{\mbox{\scriptsize 0.057}}$	$31.9_{\mbox{\scriptsize 2.3}}$	$2.16_{\mbox{\scriptsize 0.30}}$	0/10
	Unfiltered Training	$46.85_{\mbox{\scriptsize 30.64}}$	$1.362_{\mbox{\scriptsize 0.649}}$	$9.38_{\mbox{\scriptsize 3.18}}$	$0.894_{\mbox{\scriptsize 0.445}}$	$65.6_{\mbox{\scriptsize 27.5}}$	$3.29_{\mbox{\scriptsize 1.11}}$	5/10
SONIC	CoFiT	$5.24_{\mbox{\scriptsize 2.76}}$	$0.750_{\mbox{\scriptsize 0.095}}$	$12.17_{\mbox{\scriptsize 3.22}}$	$0.551_{\mbox{\scriptsize 0.083}}$	$18.3_{\mbox{\scriptsize 3.7}}$	$1.05_{\mbox{\scriptsize 0.21}}$	0/10
	Unfiltered Training	$7.64_{\mbox{\scriptsize 3.37}}$	$0.801_{\mbox{\scriptsize 0.109}}$	$5.45_{\mbox{\scriptsize 0.92}}$	$0.534_{\mbox{\scriptsize 0.114}}$	$19.8_{\mbox{\scriptsize 4.1}}$	$1.34_{\mbox{\scriptsize 0.34}}$	0/10

TABLE III: Repeated hardware trials. Mean with sample SD in subscripts.

Explore similar work

Sep 14, 2026

ResSafe: Learning Safety Filtering with Residual Reinforcement Learning for Humanoids

Safe control of humanoid robots remains challenging due to their high-dimensional dynamics, contact-rich interactions, and sensitivity to disturbances. Although reinforcement learning has enabled effective locomotion and motion tracking, learned policies can still generate unsafe actions that lead to instability or falls. In this work, we propose residual reinforcement learning as an implicit safety-filtering mechanism for safe humanoid control. Instead of relying on a single nominal policy to simultaneously balance performance, safety, and robustness, we decouple performance and safety. The nominal policy focuses solely on task performance, while a residual policy learns safety corrections. This decoupling leads to a better performance--safety Pareto trade-off and avoids the need for careful tuning of multiple competing reward terms within a single policy training. We show that the residual policy can act as an implicit safety filter.

Gechen Qu, Tong Zhang, Bike Zhang +4

University of California, Berkeley · University of California, Los Angeles

Safety-Critical Control Humanoid Robot Control Robot Safety Safety Filtering Residual Policy Learning Safe RL

Ask PDF

May 29, 2026

Constrained Whole-Body Tracking for Humanoid Robots

Recent advances in reinforcement learning (RL) have demonstrated impressive whole-body agility for humanoid robots, yet ensuring safety and satisfying constraints -- particularly those specified after training -- remains a challenge. Towards this goal, we present ConstrainedMimic, a control framework that leverages whole-body kinematics and dynamics for real-time constraint enforcement within RL tracking policies. By integrating principles from operational space control and control barrier functions (CBFs), we enable the satisfaction of arbitrary runtime constraints on both the kinematic reference motion and the underlying dynamics. In whole-body motion-tracking and teleoperation experiments on a (simulated) Unitree G1 with a learned policy, we demonstrate collision avoidance (both with the robot body and external obstacles), joint limits, and center of mass stability constraints. By remaining consistent with the current contact mode and tracking objectives, we minimally restrict the capabilities of the policy when constraints are active. Our method is fully differentiable, runs on CPU, GPU, and TPU, and can be deployed at up to 300-500 Hz. All software will be freely available upon publication.

Daniel Morton, Pranit Mohnot, Marco Pavone

Stanford University · NVIDIA Research

Robotic Control Robot Trajectory Tracking Humanoid Robot Control Constrained RL Whole-Body Control Humanoid Motion Tracking

Ask PDF

Oct 8, 2026

FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems

Safe reinforcement learning commonly places safety and task performance in the same policy objective, where they can introduce competing updates. Safety filters separate them at action execution, but classical designs require an analytic safety function and dynamics model, and standard minimal-intervention filters are myopic to long-horizon task return because they minimize only instantaneous action deviation. Hard projections are also undefined when no safe action exists. We present FAITH, a feasibility-aware, model-free framework that approximates the optimal state-action safety value and amortizes minimal-intervention filtering with a feedforward network. The task policy optimizes the task return through the filtered dynamics, which recovers the feasible constrained problem without a competing safety term in the task-policy update. When no action satisfies the learned safety condition, the same filter approaches the action with minimum predicted peak harm. On a double integrator example and a Safety Gym environment, FAITH achieves the highest return among methods with no feasible-start violations and matches the lowest harm from infeasible starts. On a 29-DoF humanoid, it reaches a 99.95% safety rate while retaining 97% of the unfiltered return in Walking-Avoid, and obtains the highest measured safety rate in Push-Avoid by learning to sacrifice balancing and fall away from the protected region. The same policies are also demonstrated on a real-world Unitree G1 humanoid.

Songyuan Zhang, Baljeet Singh, Sarthak Ranjeet Kaingade +2

Massachusetts Institute of Technology · Amazon

Safe RL Constrained RL Robot Safety

Ask PDF