EmbodiWeave

EmbodiWeave / simulator

Train a policy.
Watch the robot run.

Choose a training budget and follow the robot in 2D or 3D. After training, we test the saved checkpoint on 180 episodes and check whether it qualifies for deployment.

Connecting…

New training run

Try it in the simulator

Connecting
Waiting for simulator
Step 0

The robot will appear here when training starts.

● Robot ● Pedestrian ○ Goal. Both views show the same 2D simulator.

More steps give PPO more time to learn. A longer run can still fail the deployment checks.

One run at a time. Runs stop after eight minutes.

Trainer busy or short on time? Follow the recorded judge walkthrough →

What counts as a pass?

The target test requires at least 80% success and at most 10% collisions. Three tests for earlier tasks must also pass. Finishing training saves a checkpoint even when it fails these checks.

The deployed policy remains P-static-b. The saved results show each threshold and measurement.

Choose a budget and start training.

Ask the agent

Ask about the results

Keep private information out. Your question and public graph evidence are sent to the configured model provider. The server keeps the API key private.

Model busy? Inspect the measured attempts and live graph directly →

What runs here

The agent reads FalkorDB and explains the evidence. With your training consent, it can start a 60,000-step run. Watch that run in the simulator above.

Up to 6 model turns, 10 tool calls and 2 minutes. One investigation runs at a time.

Back to simulator · Recorded Sonnet training run · Run full training locally

Tool calls

Agent activity

Ready when you are

Choose a question and agree to send it to the model provider.

    Your question and the agent’s tool calls will appear here.

    You can inspect the Cypher queries, source records and test results for each call.

    Prefer a direct query? Open the live graph console →

    About this demo

    The robot runs in a simulator. The 3D view uses the same positions as 2D; it adds no new physics. The agent uses a free model, so it may be unavailable when its provider is busy. You can still start training with the button above.