FACT: Robots Learn from Failures Using Failure-Aware Causal Training
ZeYanjie · x · 2026-08-29
World-Action Models (WAMs) are typically trained on successful demonstrations, discarding valuable data from failures. UCSD researchers introduce FACT (Failure-Aware Causal Training), a causal World-Action Model designed to learn from bad actions rather than ignoring them.
- Method: Uses an action-conditioned interface to treat failed rollouts as valid future targets. Failure data supervises the observed failed future and a lowered progress value, teaching the model consequences without forcing imitation of bad behavior.
- Architecture: A single causal diffusion transformer follows an "act, then imagine" regime.
- Results: Outperforms baselines on simulation and real-world bimanual manipulation tasks. Performance improves as more failure data is incorporated, and the model reduces success-biased hallucinations.
More from Embodied
- Huawei uses signal attenuation to detect grip hand, AI adjusts UI — aakashgupta · 2026-08-29
- Tesla automatically detects dirty windows and activates wipers — Ghost_Pilot_MD · 2026-08-29
- Luwu builds desktop embodied AI robot on Raspberry Pi 5 with $-low-cost servos — TinfoilTricorn · 2026-08-29
- Turing Award winner Sutton launches Robot Self-Evolution Academy as HICOOL wraps with 10,209 startups — 新智元 · 2026-08-29
- Japanese Official Tests Tesla, Wayve, and Waymo: Distinct Strategies in Autonomy — alexgkendall · 2026-08-29
- Delivery robot needs human help to cross the street (photo) — japie06 · 2026-08-29