Artha Labs turns existing video into simulation-ready, articulated, dynamics-calibrated scenes, drops any robot into them, and ships graded, sim-to-real-validated training data. Neutral to whose model you're training.
What the robot must do.
Video, images or text.
What counts as success.
Agents + human judges.
Sim-ready scenes + graders.
Geometry, joints, calibrated mass & friction.
DoF, reach, payload, sensors.
Any backend. Compared vs ground truth.
Hack-tested, judge-calibrated, real-eval paired.
On customer compute.
Collected, imported, or generated from human video → feeds Evaluation and Training directly.
Reconstruction is table stakes and articulation-from-video is closing fast. Artha Labs adds an agentic build loop with machine-checked critiques, and dynamics calibrated by a real-world probe.
Open Scenes →URDF / USD is solved; "what can this robot do?" is not. Import, create from video or spec sheet, and match task × scene × robot.
Open Robots →Collect (egocentric, teleop), import, or generate robot data from human video (hand inpainting + robot overlay, as in Phantom and Masquerade), gated by a quality grader.
Open Datasets →Isaac Lab, MuJoCo, ManiSkill, Newton. Backend-swapping is commodity now (Newton, Linux Foundation). The value is the sim-vs-real comparison attached to every batch.
Open Rollouts →Multi-signal graders gated on beating the free LeRobot baselines (Robometer, TOPReward) on an adversarial suite. Real-world eval reported with confidence intervals.
Open Evaluation →openpi, Isaac-GR00T, LeRobot stacks. Graders as reward, real demos co-trained, jobs dispatched to the customer's cluster.
Open Training →Mock only — illustrative data. Shared stores across modules: scene · robot · dataset · grader.