robotics
One World Model Learned a 28-Dimensional Robot Language
Pelican-Sim trained on about one million trajectories and distilled 35 rollout steps to four for a reported 5.67-fold speedup.
Summary
Pelican-Sim trained on about one million trajectories and distilled 35 rollout steps to four for a reported 5.67-fold speedup.
Pelican-Sim 1.0 predicts future observations from images and robot actions across heterogeneous bodies. Its shared 28-dimensional action space is paired with rendered action videos that connect robot descriptions and camera views to pixels. Sparse mixture-of-experts layers absorb differing dynamics, while causal adaptation and distillation reduce autoregressive rollout generation from 35 steps to four. The authors report a 5.67× speedup and stronger controllability and video-quality metrics on AgiBotWorld Beta, RoboMIND and RoboTwin after training on roughly one million real and simulated trajectories. Those are simulator benchmarks, not evidence of general physical-world reliability.
Why it matters
Pelican-Sim trained on about one million trajectories and distilled 35 rollout steps to four for a reported 5.67-fold speedup.
Limits and context
- Those are simulator benchmarks, not evidence of general physical-world reliability.
Key claims
Pelican-Sim trained on about one million trajectories and distilled 35 rollout steps to four for a reported 5.67-fold speedup.
Qualification: Those are simulator benchmarks, not evidence of general physical-world reliability.
Evidence: source-2026-09-14-008
Sources
- arXiv preprint 2609.12036arXiv · primary research
Corrections
No corrections have been recorded for this story.