AffordanceWAM Uses Human-Video Affordances to Transfer Robot Manipulation Without Action Labels
PKU · hf · 2026-10-20
Researchers from HKU, PKU, HKUST(GZ) and others propose AffordanceWAM, an affordance-aware generative World Action Model built on a pretrained video diffusion Transformer. Separately parameterized World and Action Experts, coupled via Masked Joint Self-Attention, jointly predict future RGB observations, scalar affordance fields, affordance heatmaps and continuous robot actions under a unified flow-matching objective. Human videos supervise the world streams while robot trajectories provide action supervision only, enabling transfer without human action labels, with consistent gains on RoboCasa, CALVIN ABC→D and real-world manipulation.
More from Embodied
- New "Embodied Olympics" Benchmark: GPT and Claude Agents Play Soccer and Ping Pong — ManlingLi_ · 2026-10-20
- SF founders go deep in Shenzhen's Huaqiangbei building BCIs and wearables — LoganGrasby · 2026-10-20
- MicroAGI trains a handyman model with tens of thousands of hours of video — isnit0 · 2026-10-20
- Uber-linked group secretly campaigns to ban robotaxis in Washington D.C. — gabriberton · 2026-10-20
- Meta admits smart glasses were recording conversations without the wake word — Resident-Fall4351 · 2026-10-20
- Wayve demos its Fiat 500e autonomous prototype on Turin's iconic Lingotto rooftop track — alexgkendall · 2026-10-20