Robotics simulation & evaluation
Simulations that
match your robot.
Describe an environment and Brio builds it in physics. Point it at your policy and it runs — every episode on video, scored, in a sandbox you never have to provision.
Episode
3 / 10
SPL
0.74
Step rate
512hz
Distance
11.4m
Contacts
6
mujoco · rby1 · seed 42 · sandbox brio-a71d0935
The console
Click through the whole loop.
Describe an environment, sample the tasks you want to evaluate against, point a scenario at your policy weights, and simulate. This is the real flow, with sample data — pick a step and drive it yourself.
a modern open-plan office with desks, chairs, and a meeting area
scene.glb
Open-plan office
Generation
Chat
Environment
Retrieved assets
Retrieved from a 131k-asset index — nothing is invented, so what you see is what the physics engine loads.
What Brio does today
Environments on demand, evaluations you can trust.
Brio generates simulated environments from a prompt and runs your policy inside them on managed GPU sandboxes. No rigs to build, no cluster to operate, no eval harness to maintain.
Worlds from a sentence
Warehouses, hospital wards, data centers, apartments. A prompt becomes a laid-out scene with real assets and collision geometry, exported as MuJoCo physics and a web preview — usually in minutes.
Seeded, reproducible tasks
Start pose in one room, goal in another, planned reference path. Same seed, same episode — so a regression is a regression, not a dice roll.
Bring your own policy
OpenVLA, SmolVLA, π0.5, or a classical planner stack. Weights come from your Hugging Face or GitHub account over OAuth, pulled at run time.
MuJoCo and Isaac Lab
Pick the simulator that fits the question. Each attempt gets its own container, its own queue, and the same result shape, so you compare across backends without rewriting your harness.
Your weights stay yours
No credential ever enters a sandbox, artifact tokens are write-only, logs are redacted before they leave the container, and a sandbox is destroyed the moment its attempt ends.
Evidence, not a number
Video of each episode, pass/fail, path efficiency, the per-episode breakdown, the full sandbox log, and the GPU wall-clock it cost. When a policy fails, you can watch it fail.
Where we are going
A simulation that matches your robot, not a guess at it.
Generated environments close the cost of building worlds. They do not yet close the sim-to-real gap — that takes a world model fit to the reality your robot operates in. That is what we are building toward, and everything shipping today is a step on that path.
Fit to your sensors
RGB, depth, lidar, and proprioception rendered the way your stack receives them — including the noise, latency, and dropout your hand-built sim leaves out.
Learned from your logs
Point Brio at trajectories you have already collected, and the environment gets fit to the contact and friction your robot actually meets, instead of a guess at them.
Branch reality
Take the run that went wrong and fork it into a thousand variations. The rare events that decide real deployments become part of the eval set, not a postmortem.
Team
Built by engineers who have shipped robots, not slideware.
We spent years watching policies that aced simulation fall apart on real hardware. Brio is the tool we wanted in those rooms.
Rohit Gore
Co-Founder & CEO
Patrick Walsh
Co-Founder & CTO
Grab a time on our calendar
Get access
Stop building worlds. Start scoring policies.
We are onboarding robotics teams in waves — bring a policy and a prompt, and we will run the first evaluation with you.
Start now
Sign up