SimEX lets coding agents master real robot skills via sim-integrated autoresearch
yifengzhu_ut · x · 2026-10-06
SimEX (arXiv:2609.38982), from UT Austin-affiliated researchers incl. Peter Stone and Rocky Duan, tackles coding agents' weakness in the physical world. It runs two stages: open-ended probe-and-optimize iterations in simulation to build a robust robot toolbox, then a few real-world trials that jointly correct the toolbox and simulator, using the fixed simulator to diagnose failures and screen repairs. On towel folding, barcode scanning, and plate manipulation, agents acquire skills with zero demonstrations.
More from Embodied
- ProgressCompass: context injection cuts embodied progress reward model error by 63% — Jianshu Zhang · 2026-10-07
- Video2Skill benchmark: most of 19 open-source VLMs fail at streaming embodied skill discovery — Jianshu Zhang · 2026-10-07
- MM-ABC robot foundation model hits 83% on real-world mobile manipulation tasks — Qiwei Liang · 2026-10-07
- Cortex Harness lets frontier models run real robots, with human takeovers feeding training data — kimmonismus · 2026-10-07
- 60 builders to voice-enable a Furby at SF Tech Week hardware hackathon — AssemblyAI · 2026-10-07
- Browser-based Bluetooth support lets you teleoperate robots with a PS gamepad — chrismatthieu · 2026-10-07