Arithmancy (@ArithmancyAI) header
Arithmancy (@ArithmancyAI) profile picture

Embodied AI needs messier data. Physics validated edge cases for robots before they meet us.

San Francisco, CA
Joined April 2026
Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)Photo by Arithmancy (@ArithmancyAI)
We’ve built and proven: Cross-engine divergence detection — replay an identical control stream through multiple engines, measure where they disagree. Receipt: 68.7 mm (MuJoCo holds, MuJoCo-Warp never lifts). Reproducibility / integrity verification — bit-identical repeatability within an engine, byte-identical regeneration from seed, command-stream integrity 100/100. Receipt: GPU test — byte-identical across 66 pairwise comparisons, 0.000 mm. Collision-configuration detection — flag when the compiled model can't compute the contacts its metadata describes. Receipt: 92/100 → 0/100. Our current plans: Sim-vs-hardware calibration delta — measure where a customer's sim diverges from their actual hardware. Receipts: jaw 14.3 vs 12.5 mm, camera 47.2° vs 70.7°, arm sag on commanded height.
1
1
240
This week we had our first call with a partner for a potential pilot Thanks to @boardyai
1
95
We opened up our Mujoco simulation data pipeline platform Free, no sign up required. Describe a scene in plain language. The pipeline builds it in MuJoCo, runs the physics, and packages verified episodes into training-ready datasets. Asimov compatible, physics accurate. Link in comments
1
1
119
We ported a working MuJoCo task to Warp and three things broke silently. Byte-identical does not mean physics-preserving. (notes from a migration, with numbers.) Most robotics teams are about to move simulations they've trusted for years onto a new GPU-native stack: Newton, MuJoCo-Warp, or both. We did it on a small pick-and-place task and kept a record of every run. The task worked on CPU MuJoCo. After the port it never lifted the object. Nothing errored. Every run was reproducible. It was just wrong. Here's what we found, in the order we found it. 1. The port dropped every actuator filter. Our arm had nine actuators with first-order filters on their dynamics. After conversion to Warp, all nine were gone. No warning, no log line. The model compiled, ran, and produced smooth, deterministic, physically different motion. The gripper closed with the wrong profile and the grasp failed on every one of 12 replays. If you only compare trajectories, you'll see a divergence and blame the engine. The engine was fine. The model it received was a different robot. 2. The object's inertia changed. The source model declared an explicit inertia for the object being picked. The target recomputed it from geometry. The mass was the same; the tensor wasn't. For a 15.8 mm brick that's the difference between a stable grasp and a rotate-and-drop. This is the one that worries us most, because it's invisible in the file. You have to compare compiled models, not XML. 3. The metadata described contacts the compiled model couldn't compute. Our collision masks, as recorded, described contact pairs that the compiled model had filtered out. Once we forced the compile to honor the recorded masks, lift success went from 92/100 to 0/100, meaning the "working" version had been succeeding on physics it wasn't actually simulating. We'd been passing our own test on a bug. What we ruled out GPU nondeterminism. 12 replays on two RTX 3090s, 66 pairwise comparisons, byte-identical state, max divergence 0.000 mm. If your migration looks flaky, look at the model before you look at the hardware. (Stack: MuJoCo/Warp 3.11.0, Newton 1.5.1, Warp 1.16.0.) Why this generalizes None of these are Warp bugs or MuJoCo bugs. They're the class of defect that lives in any port between engines: the semantics of actuators, inertia, and collision filtering differ across engines, and converters make reasonable default choices that happen to change your robot. Reproducibility tests can't catch them because the wrong model is reproducibly wrong. What catches them is comparing compiled models on the properties that determine contact-rich behavior, and doing it on every commit, not once during the migration. What we're releasing We’re scoping an open-source checker to catch this. It diffs source and target models after compile across actuators, inertials, collision configuration, solver settings, joint properties, and sensors, and fails your CI on breaking changes. It doesn't run physics and doesn't tell you which engine is right. It tells you what your port changed. If you're mid-migration and want to know where your engines actually disagree on your task, send us a model and a control stream. We're running divergence audits free while we validate the method - DMs open.
1
1
1
285
We started out building a pipeline to generate training data that's hard to collect in the real world. It worked, up to a point. You describe a scene, our pipeline renders it in MuJoCo, and you get simulation data you can send to your robot. But when we tested it, the robot's ability to perceive distance and grip still fell short. Rendering a scene and simulating the physics wasn't enough. So we asked why, and we started comparing our simulation directly to our real robot. That's when we found it. The gripper was off. The camera was degrees off. Nobody had caught either one until after we deployed. Then we realized something bigger: that step, checking whether the sim actually matches the robot, is missing from how simulations are verified before they're deployed. Almost nobody does it. And this is our pivot, Arithmancy is now focused on validation.
1
1
2
192
We're keeping the simulation pipeline running as a platform people can use to build training data. It can still be accessed at sim.arithmancy.ai
1
29
Announcement coming soon
2
119
We're opening the pipeline to a small number of design partners — teams with a real robot, a real failure mode, and a metric they already track. Not a signup list. A handful of teams we build with. If you can name one failure mode that costs you money, let's talk.
60
The single most important trend in robotics right now is the shift from VLA models to World Action Models. NVIDIA's blog: "WAMs were a small subfield within VLA research."
1
115
All four share one dependency: action-conditioned trajectories with real dynamics.
1
16
That's what we generate. Physics-gated state-action data with ground-truth forces and deterministic replay.
15
We deliberately degraded the camera to stress-test the policy. It still held 70%. Three attempts to close that gap through training failed. We documented them as failures. Our ledger shows all 46 of 76 runs — passes and failures side by side.
67
We trained a world model on our own generated data. It tracks the object to 0.8 cm, closed loop — and that state is what the policy sees. Not a bigger model. Better data: 5,000 physics-gated episodes, generated in one command. #worldmodel #sim2real #robotics
73
The catch, because it matters: All of that is in simulation. The first physical pick on our real arm just landed — via the coordinate-fed fallback, not the camera policy. The camera policy has only seen rendered images. Calibration's next. We'd rather say that than pretend.
1
83