Artha Sim
One API over every backend. We don't build physics.
Wire kitchen_counter_v3 × Humanoid-G1 and fan out millions of rollouts. Pick the engine that fits the task — Artha handles the translation, domain randomization, and rollout capture.
Deliberately not a moat. Building another physics engine is wasted capital — Isaac Lab, MuJoCo MJX, and ManiSkill already do GPU-parallel sim. vs. Genesis's viral "43M FPS": that was a near-degenerate config (1 substep, self-collisions off, idle robot); realistic settings drop it ~150×. We report honest, task-matched throughput, not headline numbers.
Backend
auto-selected for task: contact-rich manipulation
⚡ MuJoCo MJX ~0.9M steps/s
🎨 Isaac Lab (RTX) vision-in-loop
🧪 ManiSkill 3 ~30k FPS
🌀 Genesis multi-physics
MJX chosen: superior contact model for grasping; switch to Isaac Lab when photoreal vision dominates.
Rollout viewer
4,096 parallel envsDR: visual=high physics=low
env 0 / 4096 · task: load_dishwasher · step 312/600
success: 71.4% · ▲ 2.1%/epoch
Recent rollout batches
| Batch | World × Robot | Policy | Backend | Success | Real-eval Δ | |
|---|
| #5512 | kitchen_v3 × G1 | π0.5 (ft) | MJX |
71% |
−6% (honest) |
grade → |
| #5510 | kitchen_v3 × YAM | GR00T N1.7 | MJX |
44% |
−19% gap |
grade → |
| #5501 | warehouse_v1 × G1 | MolmoAct 2 | Isaac Lab |
83% |
−4% |
grade → |
The "Real-eval Δ" column is the product. Every sim number is paired with the measured gap on real hardware (RoboArena-style). A 19% gap is surfaced, not hidden — because data that aces sim and fails on contact is worthless. No competitor leads with this.
Mock only — illustrative data. See Artha.md §4c & §7 (sim-to-real cliff).