Skip to main content

Policy evaluation on the YAM station

SE3 Labs evaluates your robot policy on a real bimanual YAM station without your model code leaving your machine. Your policy runs wherever you like. The station streams observations to it, executes the actions it returns under the station's safety layer, and records everything. An operator on site supervises, scores each episode, and resets the scene.

This site documents what you need to connect: the client library, the wire protocol, what an observation and an action contain, and what comes back.

Getting started

Install the client, run a reference policy against your job, then wrap your own model.

Protocol

The gRPC stream: handshake, negotiation, episode messages, and the rules the station enforces.

Observations & actions

Every field the station sends you and every field it accepts back.

Results

Episode outcomes, invalidation reasons, and the job summary you receive.

How it fits together​

  • You dial in. The station never connects out to you, so nothing on your side needs to be reachable from the internet.
  • The job id is the credential. The operator gives you a random, unguessable id for each job. Presenting it attaches your policy to that job. Treat it like a password.
  • One stream per job. The whole session, from handshake to the last episode, is one bidirectional gRPC stream.
  • The station enforces limits. Joint step and speed caps, a workspace box, a collision check, and a deadline per step. Anything outside them is clamped or rejected and counted, never executed.

The session in brief​

  1. You connect with the job id and a manifest declaring your model, the action schema you speak, and the camera streams you want.
  2. The station replies with what it will actually run with: image sizes, step rate, chunk horizon, deadline.
  3. A short preflight measures your latency and validates a few chunks without moving anything.
  4. For each episode: a Reset, then observations at the negotiated rate, each answered with a chunk of actions. Your policy may end an episode early by sending EndEpisode. The operator resets the scene between episodes and scores each one.
  5. Close ends the session. You receive a summary with the success rate over scored episodes and the per-episode outcomes.

Start with getting started.