BEACON paper accepted to NeurIPS 2026: learns per-sample trust weights for cross-domain robot policy co-training
hankyang94 · x · 2026-10-27
A paper from Heng Yang's team (led by Antong Zhang and Han Qi), BEACON, is accepted to NeurIPS 2026 on mixing cross-domain data for robot policy training.
- Key idea: instead of fixed-ratio dataset mixing, learn how much to trust each source sample, casting co-training as discrepancy-aware importance reweighting guided by the Best-Effort Adaptation (BEA) bound
- Jointly trains a diffusion visuomotor policy and per-sample source weights, with scalable instance-level discrepancy proxies, stochastic alternating updates, and a multi-source extension
- Across sim-to-sim, sim-to-real, and multi-source manipulation, BEACON beats target-only, fixed-ratio, and feature-alignment baselines on robustness and data efficiency, and implicitly aligns representations without an explicit objective
More from Embodied
- Toyota Pours $6.4B a Year Into Humanoid Robots, Targets 400,000 by 2028 — emmanuelvivier · 2026-10-27
- GPT-6 moves a camera to insert a flower, yet burns an egg: reflections on the Astra report — liuziwei7 · 2026-10-27
- Harness Device ditches paper box for premium UV-printed metal tin packaging — dee_hw · 2026-10-27
- $299 Desktop AI Assistant Intern 2 Adds iMessage, Ships Preloaded With OpenClaw Agent — dee_hw · 2026-10-27
- UBTECH opens humanoid robot factory producing one robot every 10 minutes — import_jmr · 2026-10-27
- Nebius × Positronic Physical AI Competition opens: top 10 sim policies run on real robot arm — lukas_m_ziegler · 2026-10-27