UBTech wants its robots to learn in a copy of the real world. The Chinese humanoid maker launched Thinker-Sim on Nov. 2, a simulation platform built around what it calls a Real2Sim2Real loop: digitise real places, train and test robots in simulation, then validate on physical machines.
How it works
- Scene reconstruction: UBTech's Thinker-GS algorithm and automated SimReady asset generation map real factories and showrooms into 3D, then add robots, equipment and people for each task.
- Synthetic data: UBTech says its own rendering engine is 67.3% faster, a single eight-GPU node produces about 390 hours of synthetic data a day, and custom SIL/HIL communication modules raise efficiency by more than 400%.
- Benchmarks: a unified test suite covers perception, navigation and manipulation, randomising lighting, textures and materials to test robustness.
- No-code workflows: one-click templates and visual workflows cover over 90% of target scenarios, and an Ulab-Agent lets users manage pipelines in natural language.
The catch
Gasgoo points out that synthetic data still has to be proven on physical robots, and how narrow the sim-to-real gap becomes will depend on real deployments. Collecting data robot by robot is slow and costly, which is why UBTech, Baidu and Tesla, which trains Optimus partly from video and simulation, are all leaning on simulated worlds.
