D&A turns any robot task video into an off-the-shelf reward model with zero training
KyleMorgenstein · x · 2026-11-04
The D&A method takes a (partial or full) video of any robot performing any task plus a text description and produces an off-the-shelf reward model with no training. Demonstrated on the SGS locomotion task; notably, D&A only grants extra reward when the goal is visible in the image, showing its reward signal aligns with the visual goal.
Related event: D&A Turns Any Robot Video into a Ready Reward Model Without Training(2 posts)→
More from Embodied
- Chore-Doing Lamp Lume Raises $11M Pre-Seed, Named TIME Best Invention — tkexpress11 · 2026-11-04
- Economist Tabarrok: 95% of My Tesla Miles Are FSD, It's Redefining Car Ownership — pduan · 2026-11-04
- Robotics team preps customer-site live demo with only 1.5 days to collect data and train — DominiqueCAPaul · 2026-11-04
- LeRobot datasets now replay in 3D with OpenUSD robots: SO-101, OpenArm, Unitree G1 — mishig25 · 2026-11-04
- Screen-free AI storytelling toy storymaker raises $2M, ships mid-November at $89 — Scobleizer · 2026-11-04
- Robotics week in review: Tesla Optimus production line, PI at ~$20B valuation — lukas_m_ziegler · 2026-11-04