UCLA team unveils OpenSLA, a sensor-language-action model trained on 79 modalities and 116K individuals
yang_yuzhe · x · 2026-10-08
A UCLA team (led by Yuekai Xu and Zitao Shuai, with Yuzhe Yang) releases OpenSLA, a Sensor-Language-Action foundation model that jointly learns actions and language grounded in real-world sensor observations, rather than training them separately.
Three core capabilities:
- Predict: multi-level action learning from sensor data (necessity, category, fine-grained label)
- Understand: recovers physiological states and fine-grained measurements (heart rate, glucose, respiration) from raw signals
- Explain: grounds each action in the sensor evidence and individual context supporting it
The framework spans 7 datasets, 79 sensor modalities, 116K individuals, and 60 action groups, from clinical care (zero-shot transfer to the external MIMIC-IV cohort) to everyday CGM glucose monitoring. The model also supports zero-shot prediction of unseen actions and readouts of future physiological states. Paper, code, and models are open-sourced.
Related event: UCLA Team Open-Sources OpenSLA: Unifying Sensors, Language, and Action(7 posts)→
More from Embodied
- Learned latent dynamics roll forward from few observations to recover hopper state with uncertainty — hisspikeness · 2026-10-08
- Adding a pixel-reconstruction loss makes latents unable to decode hopper state — hisspikeness · 2026-10-08
- ASUS partners with Nous Research to ship Hermes Agent on RTX Spark PCs — Teknium · 2026-10-08
- RoboWorld Challenge 2026 opens: embodied world modeling across three tracks — liuziwei7 · 2026-10-08
- E2E self-driving startup Wayve lands Stellantis deal after Nissan, claims it beat Tesla to E2E — alexgkendall · 2026-10-08
- Rabbit OS3 update lets the R1 place outbound phone calls — SimonBalmain · 2026-10-08