RoboSimStudio Turns One Phone Video Into an Editable Robot Physics Twin in Under an Hour
lukas_m_ziegler · x · 2026-10-24
RoboSimStudio, an open toolkit built on Newton by UC Berkeley and NVIDIA Robotics researchers, showed a striking demo:
- Input is just a smartphone video — no CAD, no calibrated cameras. AI agents built a physics twin of a bimanual task (one arm holding a bar, the other wiping a marker stripe off with a sponge); Claude and GPT-6 Astra each built one independently, the fastest in 56 minutes.
- Closed sim2real loop: on the real robot the bar's support actually moved (modeled as fixed in sim, so the sponge lost contact). The agent went back, replaced the fixed constraint with a free rigid body, and fixed the simulation.
- Every physics parameter is exposed as a human-slidable slider or agent-editable field.
The video-in, sim-out, real-robot-verified, agent-self-correcting pipeline hints at a new paradigm for sim2real training.
More from Embodied
- Dev rides 1,000km of Chinese high-speed rail, snaps Shenzhen's police patrol robot — LoganGrasby · 2026-10-24
- Niantic Spatial CEO Inhi Suh on sim-to-real infrastructure for embodied AI — ExtensionEcho3 · 2026-10-24
- WorldOracle benchmark from Princeton, NVIDIA, AgiBot exposes hallucinations in action-driven world models — jiqizhixin · 2026-10-24
- Waymo robotaxis are now doing In-N-Out drive-thru runs fully on their own — FanaHOVA · 2026-10-24
- SenseTime open-sources 27B robot model matching GPT-6-Astra at ~1/50 the cost — rohanpaul_ai · 2026-10-24
- Samsung TVs found running facial recognition code that returns face bounding boxes — basedjensen · 2026-10-24