Samsung and SJTU unveil RoboICL: frozen GPT-6 Astra learns robot tasks via in-context learning
jiqizhixin · x · 2026-10-09
Samsung and Shanghai Jiao Tong University present RoboICL (Embodied In-Context Learning with GPT-6 Astra), an alternative to training task-specific VLA models: a fully frozen frontier model that learns to act just by watching.
- Core idea: no parameter training and no task-specific VLA. At inference time, the model receives a few demonstrations plus a running record of what the robot just did and how the environment changed.
- How it works: demonstrations and execution history are organized into embodied context the frozen GPT-6 Astra consults at every decision step, directly outputting the next dual-arm action.
- Capabilities: precision insertion, long-horizon dual-arm coordination, and operations that depend on the initial state.
It effectively ports the LLM in-context learning paradigm to embodied control, with potential implications for the VLA training route.
More from Embodied
- Robotics insiders say humanoids are a bubble; only Unitree has real revenue while Figure sits at $39B valuation — cixliv · 2026-10-09
- SatNav: Scalable Long-Horizon UAV Vision-Language Navigation Benchmark From Satellite Imagery — Jiajun Jiang · 2026-10-09
- Will Brown: robotics isn't scaling because researchers haven't learned LLMs' subtle art of scaling — willcb · 2026-10-09
- Shenzhen OEM smart rings cost $60 to make — white-label hardware has no moat — LinusEkenstam · 2026-10-09
- WHOOP keeps logging easy runs as Zone 5, user says Apple Watch gets it right — marclou · 2026-10-09
- China Begins Deploying Fleets of Self-Driving Delivery Robots for Overnight Operation — Polymarket · 2026-10-09