Artha
product mock · v0.1
The data & eval layer for Physical AI

Video in → physically-faithful worlds, graded rollouts, RL-ready data.

Artha turns the world's existing video into simulatable, articulated, collision-correct environments, drops any robot into them, and ships auto-graded, sim-to-real-validated training data. Neutral to whose brain you're training — the picks & shovels, not another humanoid.

Build a world from video → See the RL loop
The loop — bringing "rollout → grade → reinforce" to robots
01 · World Store

Video → World

Phone clip → 3D geometry, objects, joints, mass & friction.

02 · Robot Store

Capability model

DoF, reach, payload, sensors → drop-in agent.

03 · Artha Sim

Rollouts at scale

Backend-agnostic GPU sim. Millions of episodes.

04 · Graders

Grade & reinforce

Score, reward, one RL epoch — validated on real hardware.

Four products, one platform

World Store

From passive video to a physically-correct, articulated, simulatable world.

Reconstruction (3DGS/NeRF) is table stakes. We add the layer no one ships: inferred joints, mass, friction, compliance + visual and physical consistency. Take a World Labs/Marble or Cosmos export and make it actually simulate.

Open World Store →

Robot Store

A queryable capability representation for any embodiment.

URDF/USD is solved; "what can this robot do?" is not. Learn DoF, reach, payload, dexterity and sensing from a few videos/specs, and reuse across embodiments.

Open Robot Store →

Artha Sim

Auto-wire world + robot into the right backend. Don't rebuild physics.

One API over Isaac Lab, MuJoCo MJX, ManiSkill, Genesis. Value is orchestration + domain randomization, not a new engine.

Open Artha Sim →

Graders & RL

Write graders, score rollouts, close the loop — without a sim-to-real cliff.

An open Grader Store with video-grounded, non-hackable success detection, reward shaping, and RoboArena-style real-world eval as a first-class product.

Open Graders & RL →