Zero-WAM: Robots Learn New Tasks Zero-Shot from Human Video Prompts

A team led by researchers including Junwei Liang at HKUST(Guangzhou), publishing as Robbyant, released Zero-WAM, which brings in-context learning to robotics: a human manipulation video serves as the prompt, and a causal video-action model, trained with future-block prediction objectives, directly generates robot actions—enabling zero-shot execution of tasks never seen in training, without task-specific fine-tuning.

Confirmed

• The method is framed as In-Context World-Action Modeling, conditioning on a human video context (original post by @Robbyant-Research)
• A HumanGen data pipeline was built for scale, auto-generating 74.2K human-robot ICL pairs across 8.6K tasks
• Anti-cheating mechanisms were used during training to ensure genuine in-context learning rather than memorization (noted by @CyberRobooo)
• Only human demonstration videos are required, not robot-specific demonstration data

Why it matters

• Human videos are far easier to obtain than robot demonstrations; using video as a prompt sidesteps the per-task fine-tuning data bottleneck, a point emphasized by reposters such as @deepakpathak and @ZeYanjie
• Transferring the LLM-style in-context learning paradigm to embodied agents could change how robot tasks are programmed, pending further real-robot validation

2026-08-28 ~ 2026-08-29 · 5 related posts

• [source] Zero-WAM: In-Context World-Action Modeling for Robotic Task Generalization — Robbyant-Research · 2026-08-28
• Zero-WAM: robots learn unseen tasks from a single human demo video, no fine-tuning — deepakpathak · 2026-08-28
• Zero-WAM lets robots learn unseen tasks in-context from a single human video, hitting 46.95% zero-shot success — ZeYanjie · 2026-08-29
• [source] Robbyant Introduces Zero-WAM: Robots Learn Tasks from Human Video Demos via ICL — CyberRobooo · 2026-08-29
• [source] Zero-WAM Lets Robots Learn Tasks from Human Videos — CyberRobooo · 2026-08-29