Robots
A capability model for any robot.
URDF/USD tells you the geometry. Artha Labs' capability layer answers the question that actually matters for training: what can this robot do — reach, payload, dexterity, sensing — and how does that transfer across bodies.
vs. Menagerie / Isaac asset zoos: those ship kinematics & dynamics. We add a queryable capability abstraction and cross-embodiment transfer — both unstandardized, both open research, both reusable across every scene.
IMPORT
URDF · USD · MJCF
Validated on import: joint limits, inertia tensors positive-definite, collision meshes present.
CREATE
From video or spec sheet
When no model exists: infer kinematics and envelopes (below).
EVALUATE
What can this robot do?
Capability card + task-fit query against a scene.
Registry
🦾
Franka Panda · DROID
tabletop · 7-DoF + gripper
used in our lab runsDROID 76K traj
Reach
Payload
Dexterity
Drop into sim →
🦾
SO-ARM-101
tabletop · 5-DoF + gripper
open / cheapLeRobot
Reach
Payload
Dexterity
Drop into sim →
🤖
Bimanual YAM
dual-arm · tabletop · 2 × 6-DoF + grippers
hot data class720h dataset
Reach
Payload
Dexterity
Drop into sim →
🦿
Humanoid-G1
whole-body · 43-DoF · Dex3-1 hands
cross-embodiment
Reach
Payload
Dexterity
Drop into sim →
Create: learn a robot from video or spec sheet
01
Ingest a few teleop / spec videos
Watch the arm move through its range.
02
Infer DoF, limits, reachable workspace
Joint ranges, velocity/torque envelopes.
03
Estimate payload & dexterity
Grasp aperture, force ceiling, sensor suite.
04
Publish capability card → registry
Queryable, versioned, reusable across scenes.
Capability query
artha robots.query(
task = "load dishwasher",
scene = "kitchen_counter_v3",
)
# → candidates ranked by capability fit
# Humanoid-G1 fit 0.91 ✓ reach ✓ payload ✓ bimanual
# Bimanual YAM fit 0.58 ✗ reach (floor cabinet)
# SO-ARM-101 fit 0.22 ✗ reach ✗ payload
The differentiator: match task × scene × robot automatically, so a customer evaluating "which robot can do this job in this room" gets an answer before collecting a single real demo — and transfer learnings when they switch bodies.
Sensor & actuation model
| Channel | Spec | Sim model | Sim-to-real risk |
|---|
| RGB wrist cam | 2× 640×480 @ 30Hz | RTX / tiled render | low (DR'd) |
| Proprioception | joint pos/vel | exact | low |
| Tactile / force | 6-axis F/T | contact estimate | high — flag |
| Gripper | parallel, 0–85mm | actuator model | medium |
Contact-rich / tactile channels are explicitly marked high-risk → routed to real-data co-training (Training). We never pretend force sensing is solved in sim.
Mock only — illustrative data. Next: Datasets →