New post1,000 simulation environments built from real demonstrations

Closing the loop for embodied AI.

Solving real2sim2real for general-purpose robotics.

Approach

  1. 1Physics from observation

    Measure physical properties from real robot trajectories

  2. 2Real to sim

    Automatically build scene in simulation and infer correct physical parameters

  3. RL

    3Training at scale

    Scale up expert trajectories which solve the task

  4. realsimπ

    4Evaluation that predicts deployment

    Accurately predict real-world deployment results from simulation

Research & News

All posts →

Real-world data as the foundation for robotics simulations

We release 1,000 simulation environments for five bimanual tasks from the ABC dataset, created real-to-sim from real robot trajectories, and show that policies trained in them transfer to a new robot cell.

Read the post →