Lukas Ziegler (@lukas_m_ziegler)
2025-11-12 | โค๏ธ 116 | ๐ 11
Simulation as a data engine! ๐พ
@LightwheelAI is treating simulation as a data engine.
The thing thatโs so special about them is the fact they can include physics and behaviours of objects.
The companyโs pipeline measures real-world forces, mirrors them in sim, and scales learning through a human + AI + engineering loop.
Its stack is 90% automated, built on Blender human-in-the-loop QA. Every simulated contact is benchmarked against force/torque rigs and parameter sweeps, tuning friction, compliance, and material behavior.
The product line covers SIM-ready assets (rigid, articulated, deformable), synthetic data collection in VR/AR via teleoperation, and evaluation suites for closed-loop testing and reinforcement learning fine-tuning.
A team of 300+ operators gathers real and egocentric human trajectories for robotic foundation models.
Itโs like a factory for data collection! ๐คฏ
Demand is strongest from large AI labs, NVIDIA, OpenAI, ByteDance, Alibaba, looking to generate synthetic training data and RL fine-tuning runs across thousands of GPUs.
Great to see more and more ambitious people tackle the data gap problem, letโs go Steve Xie, Ph.D.! ๐ฆพ
Also a great robot demo experience at CoRL that included @LeRobotHF teleops on deformables, all in simulation!!
Good to see the combination of simulation + open-source (@ClementDelangue cc)!
~~ โป๏ธ Join the weekly robotics newsletter, and never miss any news โ ziegler.substack.com@
๋ฏธ๋์ด
