TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking
TRACE: Target-State Risk-Aligned Conditional Execution for Embodied Tracking
This research presents TRACE, a method for embodied visual tracking that balances persistence with safety. By updating only the planner head of a frozen 0.6B vision-language-action (VLA) model with proximal policy optimization (PPO), the authors improve tracking success but observe increased collisions due to continued pursuit under distraction. TRACE addresses this by introducing an execution-time risk gate that processes 84 cues and rescales translational waypoints when calibrated risk exceeds a threshold, leaving yaw unchanged. In paired evaluations, TRACE reduces the distracted-tracking collision rate by 3.96 percentage points while further improving success and following rates, demonstrating that target-state-aware execution can recover safety costs with minimal intervention.
The paper introduces TRACE, a risk-gating mechanism for embodied visual tracking that selectively intervenes on translational waypoints when collision risk exceeds a threshold, reducing distracted-tracking collisions by approximately 3.96 percentage points while maintaining high success and following rates.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2610.41275v1 Announce Type: new Abstract: Embodied visual tracking requires a robot to follow a language-specified person under occlusion, look-alike distractors, and ambiguous instructions. We show that improving tracking persistence can simultaneously increase safety violations. Using a compact 0.6B vision-language-action (VLA) tracker with a frozen backbone, we update only the planner head with proximal poli
Open source S2
date boxes, identity margins, a learned range-and-bearing probe, and the policy's commands. When the calibrated risk exceeds a threshold, TRACE rescales the translational component of the next waypoint while leaving yaw unchanged. The executed action is identical to PPO's on more than 98% of steps. In paired evaluations on identical episodes, TRACE reduces the distracted-tracking collision rate by 3.96 percentage poi
Open source S4
nts, while further improving success and following rates. Matched slowdown controls without target-state timing do not reproduce this improvement. Concurrent evaluation reproduces a collision reduction of 3.77 points, compared with 0.29 points for identical-arm null pairs, indicating that the effect exceeds evaluation variability. These results show that target-state-aware execution can recover the safety cost of pol
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org