Fudan's RoboSkill Stores Explored Code as Reusable Skills, Cutting Robot Task Time 60%
量子位 · wechat · 2026-10-24
A Fudan University paper systematically evaluates general-purpose agents directly controlling robots: with no task demonstrations or fine-tuning, GPT-6-Astra hit 100% success on LIBERO-10 within a 4-hour budget and 91.7%/80% on real-robot simple/hard tasks—but averaging 20–94 minutes per task. RoboSkill attacks the time cost two ways: tactile feedback (pressure, force/torque) lets agents verify contact states faster, cutting real-robot time from 25.35 to 15.93 minutes; and after each task the agent reviews its run and saves effective methods together with the actual executed code as Skills—dropping GPT-5.6-Sol's average from 94.5 to 35.9 minutes and lifting first-try success from 15% to 67.5%. Skills transfer across tasks and models, and keep updating through execution. Paper: arxiv.org/abs/2609.37810; code on GitHub.
More from Embodied
- On-device VLM goes production: GenieX runs Liquid AI's LFM-VL live at PyTorch Conference — PyTorch · 2026-10-26
- LeCun co-authored Temporal Straightening: straightening latent paths boosts gradient-based planning success rates — ylecun · 2026-10-26
- Humanoid robots' DRAM demand may 20x, but it's still just 0.04% of 2026 supply — zephyr_z9 · 2026-10-26
- Microduck team publishes deep dives: relocalizes within 2cm using a 64-pixel depth sensor — mishig25 · 2026-10-26
- Robot's elbow-assisted get-up move goes viral for its humanlike fluidity — iamfakhrealam · 2026-10-26
- Chinese eVTOL EHang EH216-S completes nearly 100,000 flights across 23 countries — LoganGrasby · 2026-10-26