Strike fine‑tunes 3D asset model

Strike Robot AI said on May 25 that it had fine-tuned its asset-generation model to make 3D simulation objects with dimensional ratios conditioned on real robot measurements, according to posts cited in recent social-media briefings. The company tied the update to training datasets for robot policies, where object size and geometry inside simulation can affect how learned behavior transfers to physical machines. Strike linked the work to tests on the Unitree G1 humanoid platform, a robot family already present in its public simulation tooling. ### Why does object scale matter so much in robot simulation? Robot policies are trained against distances, contact points and reachable volumes, not just pixels. A cup that is modeled too wide, a handle placed too high or a shelf opening that is a few centimeters off can change whether a grasp, insertion or avoidance maneuver succeeds in simulation and fails on hardware. NVIDIA defines simulation-ready assets as physically accurate 3D assets and digital twins that include real-world properties and physics bindings for robotics and autonomous-system training. (strikerobot.ai) That framing matches the problem Strike described: synthetic assets are useful only if their geometry and proportions line up with the robot that will act on them. ### What did Strike say it changed in the model? Strike described the update as a fine-tuning pass on its asset-generation model so that output objects preserve accurate dimensional ratios when conditioned on real robot measurements, according to the social posts referenced in the source briefings. The company said it specifically tested conditioning on the Unitree G1, which suggests the robot’s body dimensions and workspace were used as anchors for generating assets that better match manipulation and navigation constraints. (nvidia.com) Strike’s public materials place that work inside a broader embodied-AI stack. The company says its SR Platform turns real-world data into executable robot policies, while its public GitHub repository describes a browser-based Unitree G1 simulation platform built on MuJoCo WebAssembly with support for loading pre-trained control policies. ### Why use the Unitree G1 as the conditioning reference? Unitree’s own documentation presents the G1-D as part of an end-to-end humanoid data and training platform with standardized collection, model training and simulation-based evaluation. Unitree also publishes open-source tooling for G1 workflows, including simulation, reinforcement learning and robot datasets. That makes the G1 a practical reference platform for any team building simulation assets meant for humanoid policy development. (strikerobot.ai) Strike has already centered some of its public tooling on that robot. Its GitHub repository describes “Unitree Sim Web” as a browser-based simulation platform for the Unitree G1, with real-time joint control, keyframe recording and policy integration. ### How does this connect to the sim-to-real problem? The sim-to-real gap appears when a policy learns in a virtual environment whose physics, object geometry or sensing assumptions do not match the real world. (unitree.com) In manipulation tasks, geometry errors can distort grasp approach angles, collision timing and clearances. In locomotion and mobile manipulation, scale errors can also alter step planning, reach envelopes and obstacle interaction. Unitree’s open-source materials and training pages emphasize the pipeline from data collection and simulation to deployment on real machines. (github.com) Strike’s update fits inside that same pipeline: improve the physical fidelity of assets upstream so policy training data better reflects the objects a robot will encounter downstream. That is an inference from the company’s product descriptions and the social posts, not a separate formal statement from Strike. ### What comes next from here? Strike’s public footprint points to rollout through its existing simulation and training stack rather than a standalone product launch. The company’s website lists the SR Platform as its embodied-AI training stack, and its GitHub repository shows active work on Unitree G1 simulation infrastructure. The most concrete next checkpoints are whether Strike publishes asset examples, benchmark results or broader robot-platform support through those channels. (unitree.com) (strikerobot.ai)