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
- Chinese startup DexForce trains humanoid robots with virtual data to slash costs — pstAsiatech · 2026-10-15
- EmbodiedEvo tops RoboDojo at 38.95% success by letting robots inherit failed practice — thetripathi58 · 2026-10-15
- VR dev's one line stuns a Quest 3 owner who played 3 games for 12 months ignoring mixed reality — Aiden_Tech_Ai · 2026-10-15
- ROBIB robotics meetup #3 lands in Ghent on Nov 19 with academic and industry speakers — MarwaEldiwiny · 2026-10-15
- How 3D Gaussian Splatting gives robots spatial memory of what's out of view — TheTuringPost · 2026-10-15
- RTX Spark closes the gap with M5 Macs, ushering in an "age of preference" for PC buyers — BenBajarin · 2026-10-15