CMU's SeeQ: generalist robot value function on a VLM nearly doubles real-world task success
aviral_kumar2 · x · 2026-09-29
CMU's Aviral Kumar team introduces SeeQ (Subtask-elicited Q-functions), a generalist language-conditioned value function built on a pretrained PaliGemma VLM. SeeQ autoregressively predicts the active subtask in natural language before estimating its Q-value, shortening the TD credit-assignment horizon and removing the need for humans to specify subtasks at deployment. Pretrained on open robot manipulation data and used for best-of-N action selection, it nearly doubles success rates across four real-world long-horizon bimanual tasks.
Related event: CMU's SeeQ: Generalist Robot Value Functions Nearly Double Success Rates(3 posts)→
More from Embodied
- Walden Robotics says good evals are gold: judging robot policies on autonomy, interventions and transfer — adnothing · 2026-09-29
- User pays 30% premium over Hertz just to rent a Tesla with FSD — a clear PMF signal — yunta_tsai · 2026-09-29
- Agility Robotics shows off dexterous manipulation of varied objects at IROS — lukas_m_ziegler · 2026-09-29
- Uni-VLaT adds touch to VLA policies: back-tap walking goes from 0% to 85% success — stepjamUK · 2026-09-29
- First Microducks robotic ducks being assembled in Bordeaux for testing and demos — andimarafioti · 2026-09-29
- Desk robot Lamp sniffles after being called a "desk accessory" on Eleven v3 — dee_hw · 2026-09-29