The authors propose WATCH (World-Action model Training from Cross-Human videos), which pre-trains a video-action model using pairs of human demonstrations of the same task from different environments, viewpoints, and demonstrators.
Cross-video human demonstration pairs improve robot world-action model transfer
WATCH pre-trains world-action models by predicting one human demonstration from another, boosting downstream robot policy success without matched human-robot data.
Academic
Jeongeun Park · Taekyung Kim · Dongyoon Han · Sangdoo Yun
NAVER AI Lab · Georgia Institute of Technology
Research Digest··2 min read
Thread:Robot World Models
The authors introduce WATCH, a cross-video pre-training objective that learns transferable world-action representations from task-paired human demonstrations.
Why this paper
From NAVER AI Lab and Georgia Institute of Technology · Part of Robot World Models, now 27 papers
In one line
Cross-video prediction from task-paired human demonstrations learns transferable world-action representations that improve robot control.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§