Q-Planning boosts robot self-improvement from 25% to 80% success
chris_j_paxton · x · 2026-08-26
Addressing the limitation of robot foundation models on hard tasks (e.g., 25% success rate), Q-Planning introduces a learning-based harness that enables self-improvement by equipping a large visuomotor BC policy with a small off-policy Q-function.
Key Mechanisms:
- Inference: The frozen BC policy samples N candidate action chunks; the Q-function scores them, and a single-step Q-weighted average is executed.
- Training: Successful and failed rollouts are appended to a replay buffer to fine-tune only the Q-function, leaving BC weights untouched.
Results:
- On LIBERO and bimanual RoboTwin benchmarks, ten iterations lifted all scores (e.g., LIBERO-10 93% → 99%, RoboTwin 83.8% → 91.4%).
- On two contact-rich real-robot tasks, the loop improved performance purely from autonomous interaction (no human intervention) in 30 minutes.
Related event: Q-Planning Enables Robot Self-Improvement, Boosting Success from 25% to 80%(2 posts)→
More from Embodied
- Stanford's Fei-Fei Li team open-sources OpenWAM, a unified world-action model framework for robots — jiqizhixin · 2026-10-16
- Robot folding leaderboard: GPT-6 Astra leads at 1349 Elo, Muse Wonder trails at 1034 — qinzytech · 2026-10-16
- NVIDIA's TacitVLA: robots reason without reading their own thoughts, 3.4x lower latency — arankomatsuzaki · 2026-10-16
- SEAR: first large-scale controlled study of self-evolving robot agents, 7 LLMs, 330 tasks — xiye_nlp · 2026-10-16
- Podcast: subsea robotics startup Ulysses on operating 1,500m underwater and ocean carbon — blakemenezes · 2026-10-16
- Sony unveils $5,900 gold-plated Walkman with AI music enhancement — Polymarket · 2026-10-16