Neural Motion provides a generative data engine, NM Gen‑ET, that creates in‑distribution robot trajectories for any target embodiment by denoising source trajectories with a diffusion prior trained on the target robot. This lets foundation labs and fine‑tuning teams expand their training corpora without additional capture pipelines, turning data from any robot, scene, or modality into usable target‑specific data. The approach addresses the robotics data bottleneck at both pre‑training and post‑training stages.
Funding
Funding not disclosed
Founders
Product
Problem
Robotics research and development is limited by the scarcity of high-quality robot trajectory data. Collecting large datasets for each robot embodiment requires time‑consuming capture campaigns, and existing data cannot be easily reused across different robots or environments, creating bottlenecks in both pre‑training large models and fine‑tuning for specific platforms.
Solution
Neural Motion offers NM Gen‑ET, a generative data engine that synthesizes robot trajectories matching the distribution of a target embodiment. The system corrupts source trajectories from any robot, scene, or modality with noise and then denoises them using a diffusion prior trained on the target robot’s data, effectively learning and sampling from the target’s natural motion distribution. By generating in‑distribution data rather than merely retargeting kinematics, NM Gen‑ET expands training corpora without additional capture efforts, accelerating foundation model pre‑training and enabling fine‑tuning on target robots with synthetic data that mirrors real‑world behavior. This approach removes the reliance on extensive data collection pipelines and allows cross‑embodiment knowledge transfer.
Target Audience
Primary customers are robotics research labs, AI foundation model teams, and engineering groups that need large, embodiment‑specific trajectory datasets for pre‑training or fine‑tuning robot control policies.
Features
- Diffusion‑based generative model that learns the target robot’s trajectory distribution from existing data
- Noise‑corruption and denoising pipeline that converts arbitrary source trajectories into target‑embodied samples
- Supports input from any robot, scene, or sensor modality, enabling broad data reuse
- Generates native target‑distribution trajectories rather than simple kinematic retargeting
- Scales training datasets beyond the limits of physical capture campaigns for both pre‑training and fine‑tuning