RoboSSM: State-Space Models Enable Robots to Keep Improving at Test Time Without Fine-Tuning
yifengzhu_ut · x · 2026-09-28
Researchers from UT Austin (Youngju Yoo, Peter Stone, et al.) introduce RoboSSM, a scalable in-context imitation learning (ICIL) recipe built on state-space models.
- Robots learn new tasks from just a handful of demonstrations and keep improving at test time with more examples — no GPU fine-tuning required.
- It replaces Transformers with the SSM "Longhorn," gaining linear-time inference and strong long-context extrapolation, fixing Transformers' degradation on longer prompts.
- On the LIBERO benchmark, RoboSSM generalizes better to unseen and long-horizon tasks than Transformer-based ICIL, marking the first proof that SSMs are an efficient, scalable ICIL backbone.
- Accepted to IROS 2026; code is open-sourced.
More from Embodied
- EEG foundation model trained on public datasets, paired with a ~$50 BCI device — melnykowycz · 2026-09-28
- RollerCoaster Tycoon goes spatial: 1999 pixel game rebuilt with voxels on Specs — stspanho · 2026-09-28
- Robotics pioneers split on humanoid timelines: action data, not LLMs, is the bottleneck — Scobleizer · 2026-09-28
- Berkeley's TRACE robots trace monochrome data center cables via bidirectional tracing — berkeley_ai · 2026-09-28
- macOS 27 beta's new hit-test map debug tool fuels touchscreen MacBook rumors — stephancasas · 2026-09-28
- Evolutionary search on Graph-as-Policy robot controllers yields 5.27x throughput for $37 — matei_zaharia · 2026-09-28