Uranus — Exploring a Data-Driven Approach to Robot Simulation
Hi everyone,
We’re exploring a different approach to robot simulation.
Real-robot validation can be costly and difficult to reproduce, while traditional physics-based simulators often require substantial modeling effort and still leave us with the sim-to-real gap.
Uranus takes a third path: a data-driven simulation approach built around a joint-trajectory-conditioned autoregressive diffusion world model. Given multi-view observations, camera calibration, and a robot description, Uranus can generate future multi-view frames and maintain interactive closed-loop trajectories.
In our upcoming technical talk, we’ll open up the full stack behind Uranus, including:
-
3,300+ hours of real robot data — curation and calibration recovery
-
Causal training and distillation of a 1.3B-parameter world model
-
Serving infrastructure reaching 24 FPS
-
Long-horizon consistency and closed-loop policy evaluation
-
Cross-embodiment generalization across multiple robot platforms
-
Where the approach still breaks — and what we’ve learned from it
We’re also open-sourcing the code, model weights, and data so others can explore and build on the work.
This talk is the first session of D-Robotics Frontier, a technical series where we share our latest work across Embodied AI, world models, robot learning, and intelligent robotics, and open up deeper technical discussions with the broader robotics community.
October 15
20:00 GMT+8
If you’re working on robot simulation, robot learning, world models, or embodied AI, we’d be happy to have you join the discussion.
Registration: D-Robotics Frontier — #1 Uranus · Zoom · Luma