Home

Every break, labeled. Soft-body failure data for world models and robotics.

Every break, labeled.Soft-body failure data for world models and robotics.

Crash, break and failure data you cannot safely collect in the real world. OpenBeam Cloud runs a 2 kHz soft-body physics solver fully hosted, so you need no GPU, and records beam stress, breaks and contacts on every frame, synced with the video. The telemetry is yours outright, and AI agents can drive the whole loop over MCP.

3 free simulation hours. No card to start.

Every scene in this reel is soft-body physics running on OpenBeam Cloud. A Saturn V launch, a freight train, a helicopter, a race start, heavy machinery, a robot welding cell and a wrecking ball through a brick wall, each generated and filmed in the cloud with its telemetry recorded.

Failure data for world models, at scale

World models and robot policies need to see things bend, slip and break, and those are the runs nobody can film safely. OpenBeam Cloud generates them on demand as synthetic training data: soft-body physics you can vary, repeat and record. Open means open formats and data you own outright.

Run it wide

Sweep speed, angle, load and material across parallel runs. Runs are deterministic, so any sample can be regenerated exactly when you need to debug a model.

Inputs and data on one clock

Each recording keeps the video, every input and the engine data together, so you can line up what happened with what was on screen.

Straight into your pipeline

Telemetry is exported in open formats, frame-synced with every video stream, so training jobs read it directly. You own all of it, including for commercial models.

Built for AI agents

OpenBeam Cloud speaks MCP, so an AI agent can create a vehicle, robot, prop or map, upload it and run it on its own, with no person in the loop.

MCP server

Connect Claude or any MCP-compatible agent. It can create and upload assets, start runs, change parameters and read the telemetry through the same tools you use.

Hot reload

Change a model, script or map and the running simulation picks it up in place. No restart, no rebuilding the scene.

Design, test, repeat

An agent can draft a part, drop it into a load case, read what broke and try the next version, over and over, without waiting on anyone.

See it hold. See it fail.

Every clip below was recorded in OpenBeam in real time. Run experiments in parallel, one inside the design’s limits and one past it, and see exactly where it gives.

Coupe into a rigid wall at 86 km/h. Speed measured at the moment of contact.

The crumple zone does its job

The front of the car folds beam by beam and soaks up the impact. The cabin keeps its shape and the wheels stay on.

A test rig, not a viewer

Script a test, drive it by hand and record what happened, all in the same simulation.

Script your tests

Write test cases in AngelScript or Lua and run them inside the engine. Set up a load, apply it in stages and log what happens on every tick. Then run as many variations as you need in parallel.

Drive it live

Drive vehicles and operate robot arms by hand, in real time, while the solver runs.

Record and replay

Every recording stores its inputs and physics data in sync with the video. Replay a run, find the frame where something gave way, and use your recordings however you like, including commercially.

Bring your own models

Upload your own vehicles, robots, scripts and maps. Teams and Enterprise also include OpenBeam for your desktop, so you can run tests on your own machine before sending them to OpenBeam Cloud.

How hosted soft-body simulation works

The solver is the computationally expensive part of soft-body simulation, and it runs entirely on our infrastructure. The client only decodes a video stream.

Solve

The node-and-beam solver runs at a fixed 2 kHz on dedicated CPU cores, so the timestep holds even under heavy load.

Render

Each frame is drawn on a server GPU and handed straight to its hardware video encoder. It never touches disk.

Stream

Frames reach the browser over a low-latency connection, and user input returns on the same channel. No dedicated hardware is required.

Who it's for

Robot learning teams
Contact, grasp and failure data, fast: grippers that slip and parts that snap, generated in parallel with the labels attached.
World model teams
Soft-body runs recorded with their inputs and engine data, frame for frame.
AV and ADAS teams
Crash edge cases that are unsafe or too expensive to film, repeatable down to the frame.
University labs
A real solver from the laptop you already have, with runs you can regenerate exactly and cite.
Researchers and agent builders
An agent and real physics together: design something over MCP, break it, read where it failed and try again.

Pricing

Start free. Pay for simulation hours once you need more.

Plan Free betaEvery new account PersonalResearchers, students, agent builders TeamsStartup and lab teams EnterpriseCommitted annual usage
Price $0 $30 / month $150 / seat / month Custom
Simulation hours 3, then a card is required 30 a month, metered after that 100 a month per seat, metered after that Committed annual usage
What you get The full solver, telemetry export and MCP Everything in the free beta, plus private model storage A shared model library, usage by project and invoicing SSO and private deployment
Get started Join the free beta Start free, upgrade Oct 26 Book a demo Book a demo

Billing opens October 26; until then every account runs on the free beta. Founding price: the first 100 Personal subscribers who subscribe by November 1 keep $30 a month for 12 months. Hours are counted per stream while a solver is attached, so five streams for an hour use five hours; queued and idle time isn't billed. Usage past the included hours is metered, with spend alerts and caps on every account. Teams can pay by invoice.