Summary
This paper (arXiv:2607.18200) by Junyi Hu, Shuaihang Yuan, Geeta Chandra Raju Bethala, Anthony Tzes, and Yi Fang addresses robot navigation failures in cluttered indoor spaces caused by miscalibrated fixed safety margins: overly conservative margins trigger detours and timeouts, while loose margins produce near-obstacle shortcuts under perception bias. Although diffusion-based planners generate diverse trajectory candidates from egocentric RGB-D input, reliably selecting among proposals remains a bottleneck. The authors propose a context-conditioned safety critic that learns adaptive clearance preferences to rank diffusion proposals. The critic decomposes into three complementary terms: (i) a safety term with a clearance-budget penalty and control barrier function residuals for waypoint and transition safety; (ii) an efficiency term combining smoothness penalties with a safety-gated detour-ratio penalty to avoid detours without incentivizing risky shortcuts; and (iii) a distance-constraint matching term anchoring the learned budget to achieved ESDF clearance, preventing margin collapse. The critic is trained in simulation with privileged ESDF geometry and distilled into a purely perception-based selector via a two-stage teacher-student process, transferring to a Unitree G1 humanoid without task-specific tuning.
Learning Adaptive Safety Margins for Visual Navigation
Field: Robotics (cs.RO, cs.AI)
Authors: Junyi Hu, Shuaihang Yuan, Geeta Chandra Raju Bethala, Anthony Tzes, Yi Fang
Published: 2026-07-20
arXiv: 2607.18200
Overview
Robots navigating cluttered indoor spaces often fail not because they cannot generate collision-free paths, but because their fixed safety margins are miscalibrated: conservative margins lead to detours and timeouts, while loose margins cause near-boundary shortcuts under perception bias. Diffusion-based planners propose diverse trajectory candidates from egocentric RGB-D input, but reliable selection remains the bottleneck.
Method
The authors propose a context-conditioned safety critic that learns adaptive clearance preferences to rank diffusion proposals, decomposed into three complementary terms:
1. Safety term — a clearance-budget penalty plus control barrier function (CBF) residuals, ensuring waypoint and transition safety.
2. Efficiency term — a smoothness penalty combined with a safety-gated detour-ratio penalty, avoiding detours without incentivizing risky shortcuts.
3. Distance-constraint matching term — anchors the learned budget to the achieved ESDF clearance, preventing margin collapse.
The critic is trained in simulation using privileged ESDF geometry, then distilled into a purely perception-based selector through a two-stage teacher-student process.
Results
The approach transfers to a Unitree G1 humanoid without task-specific tuning.
---
*Auto-collected on 2026-07-22*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178447004