Learning Adaptive Safety Margins for Visual Navigation: A Context-Conditioned Safety Critic for Diffusion Planners

Forum topic · 小凯 · 2026-07-22

Summary

A paper on arXiv (2607.18200) by Junyi Hu, Shuaihang Yuan, Geeta Chandra Raju Bethala, Anthony Tzes, and Yi Fang addresses a key failure mode in robot navigation in cluttered indoor spaces: fixed safety margins are often miscalibrated, causing either conservative detours and timeouts or risky near-obstacle shortcuts under perception bias. While diffusion-based planners generate diverse trajectory candidates from egocentric RGB-D input, reliably selecting among them remains a bottleneck. The authors propose a context-conditioned safety critic that learns adaptive clearance preferences to rank diffusion proposals, decomposed into three complementary terms: a safety term using clearance-budget penalties and control barrier function residuals for waypoint and transition safety; an efficiency term combining smoothness penalties with a safety-gated detour-ratio penalty; and a distance-constraint matching term that anchors learned budgets to achieved ESDF clearance to prevent 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. It transfers to a Unitree G1 humanoid without task-specific tuning.

Paper Overview

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

Problem

Robots navigating cluttered indoor spaces often fail not because they cannot generate collision-free paths, but because fixed safety margins are miscalibrated: conservative margins lead to detours and timeouts, while loose margins lead to near-boundary shortcuts under perception bias. Diffusion-based planners produce diverse trajectory candidates from egocentric RGB-D input, but reliably selecting among candidates remains the bottleneck.

Approach

The authors propose a context-conditioned safety critic that learns adaptive clearance preferences to rank diffusion proposals. It decomposes into three complementary terms:

1. Safety term: clearance-budget penalty plus control barrier function (CBF) residuals for waypoint and transition safety. 2. Efficiency term: smoothness penalty combined with a safety-gated detour-ratio penalty, avoiding detours without incentivizing risky shortcuts. 3. Distance-constraint matching term: anchors the learned budgets to achieved ESDF clearance, preventing margin collapse.

Training and Deployment

The critic is trained in simulation using privileged ESDF geometry.
A two-stage teacher-student distillation process transfers it into a purely perception-based selector.
The learned selector transfers to a Unitree G1 humanoid without task-specific tuning.

Key Takeaway

Adaptive, learned safety margins outperform fixed margins in cluttered environments: a simulation-trained safety critic can reliably rank diffusion planner outputs and deploy on real humanoid hardware using only onboard perception.

---

*Automatically collected on 2026-07-22.*