Event

TRACE: A Risk-Gated Execution Method for Embodied Tracking

1 reports · 1 sources

The research team proposes TRACE, an execution-time goal-state risk-gating method: it keeps the learned policy unchanged and intervenes selectively only when risk exceeds a threshold, scaling down the translational component of the next waypoint proportionally while keeping the yaw angle unchanged. The team reports that, in paired evaluations over the same episodes, this method reduces the collision rate in distractor-tracking scenarios by 3.96 percentage points while further improving success rate and following rate. The method's tracker is a 0.6B vision-language-action model with a frozen backbone, in which only the planning head is updated using PPO. The team reports that after updating only the planning head, success rate and following rate improve across all three settings on EVT-Bench, but the collision rate in distractor-tracking scenarios rises; TRACE's risk gating was added precisely to suppress this increase. All the above effects are as reported by the authors in the paper.