Driving · 2026-10-14
Training Generalist Navigation Policies via Action Self-Distillation
Bowen Feng, Keqin Wang, Yulong Yang, Amogh Joshi, Christine Allen-Blanchette, Dhruv Shah, Felix Heide
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Action Self-Distillation (ASD) trains a language-conditioned navigation policy from trajectory-only robot datasets by using an open-source VLM as both a self-teacher and student. The self-teacher uses privileged ground-truth future trajectories to infer a language goal and structured reasoning; the student learns from onboard observations alone. Across 13 real-world indoor/outdoor tasks, the authors report improved success vs the strongest baseline.
Why it matters
The authors address an annotation bottleneck for language-conditioned reasoning/navigation by training within a loop that generates goal and reasoning supervision from trajectory feedback, without human-annotated language or reasoning labels. This enables deployable language-conditioned navigation where the teacher is absent at inference, using operator-provided natural-language goals.
Method
- Use the same pretrained VLM for a self-teacher (EMA copy of the student) and a student policy. The self-teacher receives privileged feedback: the ground-truth future trajectory projected onto the current egocentric observation.
- Teacher pass: produce an inferred natural-language navigation goal and a structured reasoning trace plus trajectory. Student pass: condition on the extracted goal (no privileged trajectory at inference) and learn to predict future trajectory tokens and reason tokens from onboard observations.
- Train with combined losses: KL divergence for reasoning tokens and cross-entropy for trajectory tokens; update teacher weights as an EMA of the student.
Limitation
The main failure mode described is execution error: the method “predicts and executes the next 8 steps at once,” so when terrain is not sufficiently flat, “the robot may deviate from the planned trajectory during execution.”
Abstract (from arXiv)
Language provides a natural interface for humans to specify and reason about goals, but training language-conditioned robotic policies and reasoning typically requires expensive manual annotations. We propose Action Self-Distillation (ASD), a method for learning language-conditioned reasoning for robotic navigation without human-annotated language or reasoning labels. ASD uses an open-source VLM, which contains broad world knowledge and reasoning capabilities, as both a teacher network and a student policy. During training, the self-teacher receives privileged trajectory feedback and infers the underlying navigation goal together with goal-oriented reasoning. The student is then trained to condition on the inferred navigation goal, predict future ego-frame trajectory, and provide a description of the scene semantics, plan, and reasoning, from onboard observations alone. Across 13 challenging indoor and outdoor real-world navigation tasks, our method improves success rate over the strongest baseline by $30\%$ in indoor settings, $30\%$ on multi-stage tasks, and achieves $95\%$ success rate on outdoor tasks, validating that privileged trajectory feedback can enable self-distillation for scalable generalist navigation policies without human annotations.
Related papers
- LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence
- DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies
- RIWANav: Recursive World-Action Models with Self-Improvement for Urban Navigation
- PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation
- Flow Policies as Actions of Skill-Level World Models: Learned and Symbolic Abstractions for Long-Horizon Planning
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC
Training Generalist Navigation Policies via Action Self-Distillation
Preprint posted: 2026-10-13 · Original record ↗
- arxiv
- 2610.17593
Explicit source versions (1)
- v1 · 2026-10-13
- Preprint available
Source date: 2026-10-13 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Language provides a natural interface for humans to specify and reason about goals, but training language-conditioned robotic policies and reasoning typically requires expensive manual annotations. We propose Action Self-Distillation (ASD), a method for learning language-conditioned reasoning for robotic navigation without human-annotated language or reasoning labels. ASD uses an open-source VLM, which contains broad world knowledge and reasoning capabilities, as both a teacher network and a student policy. During training, the self-teacher receives privileged trajectory feedback and infers the underlying navigation goal together with goal-oriented reasoning. The student is then trained to condition on the inferred navigation goal, predict future ego-frame trajectory, and provide a description of the scene semantics, plan, and reasoning, from onboard observations alone. Across 13 challenging indoor and outdoor real-world navigation tasks, our method improves success rate over the strongest baseline by $30\%$ in indoor settings, $30\%$ on multi-stage tasks, and achieves $95\%$ success rate on outdoor tasks, validating that privileged trajectory feedback can enable self-distillation for scalable generalist navigation policies without human annotations.