Rollouts
Simulate, import, compare. 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 Labs handles translation, domain randomization and rollout capture. Physical and virtual rollouts from elsewhere can be imported and compared against ground truth.
Deliberately not a moat. Isaac Lab, MuJoCo (MJX / Warp) and ManiSkill already do GPU-parallel sim, and Newton (GA Mar 2026, Linux Foundation; Disney Research + DeepMind + NVIDIA) makes solvers swappable behind one API for free. The value is in scenes, graders and the real-eval pairing, not the plumbing. 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 / Warp contact-rich
🎨 Isaac Lab (RTX) vision-in-loop
🧪 ManiSkill 3 GPU render + sim
🔀 Newton swappable solvers
🌀 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
Import & compare vs ground truth
Import rollouts
physical or virtual
🦾 Robot logs LeRobot · rosbag
🎥 Real eval episodes video + proprio
🧊 Other simulators trajectories + state
🌫 Neural sim (world model) video only
Video-only sources are scored by the VLM and rubric graders. Sim sources also get the privileged-state graders.
Sim vs real · matched episodes illustrative · batch #5512, same initial state
| Metric | Sim | Real | Δ |
|---|
| Task success | 71% | 65% | −6 pts |
| Rubric score (0–5) | 3.9 | 3.6 | −0.3 |
| End-effector path error | — | — | 2.1 cm RMS |
| Final object pose error | — | — | 1.4 cm · 6° |
| Policy ranking vs real (4 policies) | Spearman ρ | 0.8 |
Ranking agreement is the question customers care about. The reference point is RoboWorld, a neural sim that reproduced RoboArena rankings at r = 0.989. Where the physics sim ranks worse, the neural evaluator is used for that question.
Recent rollout batches
| Batch | Scene × Robot | Policy | Backend | Success | Real-eval Δ (95% CI) | |
|---|
| #5512 | kitchen_v3 × G1 | π0.5 (ft) | MJX |
71% |
−6 ± 5 pts · n=350 |
grade → |
| #5510 | kitchen_v3 × YAM | GR00T N1.7 | MJX |
44% |
−19 ± 6 pts · n=200 |
grade → |
| #5501 | warehouse_v1 × G1 | MolmoAct 2 | Isaac Lab |
83% |
−4 ± 5 pts · n=300 |
grade → |
| lab · Aug 27 | Arena home_office × DROID | π0.5 | Isaac Lab Arena |
19/27 |
measured · 9-object sweep |
|
The "Real-eval Δ" column is the product. Every sim number is paired with the measured gap on real hardware (RoboArena-style). A 19-point gap is surfaced, not hidden: data that aces sim and fails on contact is worthless. Real-eval sample size is shown because at n=40 the interval would be ±15 points.
Mock only — illustrative data. Next: Evaluation →