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
- Robocurve releases all 2,109 robot runs with transcripts and videos; Gemini 4 Argon averages 33.8% — zacharynado · 2026-10-15
- Dyson's $500 AI-powered toothbrush hit by widespread malfunctions, sales paused — Polymarket · 2026-10-15
- Autonomous Lamp: a $999 open-source desk robot lamp you can customize — dee_hw · 2026-10-15
- Inanimate preps WORKS: AI-agent-powered physical devices you control with prompts — genmon · 2026-10-15
- China's driverless L4 autonomous bus spotted operating on public roads — CurieuxExplorer · 2026-10-15
- AukiLabs Built a Real Supermarket Inside Its Office So AI Agents Can Run a Store — DominiqueCAPaul · 2026-10-15