SuperMimic trains a humanoid tracker in 3 hours on a single GPU, beating a state-of-the-art baseline that required 7 days on 128 GPUs. The approach uses differentiable planning with stable and effective gradients.
Key points
Training speed: 3 hours on a single GPU.
Technique: differentiable planning with stable and effective gradients.
Sources
- Original post
- stacx-lab.github.io - Linked in the post
Steertuning – few-shot robot task adaptation in under 5 min
Steertuning adapts a robot to a new task from five human demonstrations in under five minutes, achieving 71% success across twelve new tasks. The system fine-tunes a 400M-parameter diffusion policy for two minutes (2000 updates) and steers actions with the closest demonstration during deployment.
Key points
5 human demonstrations enable task adaptation in under 5 minutes.
71% success rate on 12 tasks in new environments.
Sources
- Original post
- Steertuning Policies from a Few Human Videos - Linked page
- github.com - Linked in the post
Fable 5.2 agents proof Fermat's Last Theorem
Fable 5.2 agents coordinated to formalize a proof of Fermat's Last Theorem within safety boundaries. The work is shared in a blog post.
Key points
Team size: 1,000 Fable 5.2 agents.
Goal: formalize a proof of Fermat's Last Theorem.
Sources
Toyota Waymo bZ4X SUV testing start 2027 expansion plans
Toyota and Waymo will begin testing self-driving bZ4X SUVs in 2027. They plan to expand the collaboration beyond Japan to the United Kingdom and other Waymo markets.
Key points
Testing self-driving bZ4X SUVs starts in 2027
Expansion planned to the United Kingdom and other Waymo markets
