robot-agent-exploration/rax/multi-env-exploration

Multi-env exploration

Exploration offers several independent environments. status lists them and their episode state; a top-level active episode is not the pool inventory. Continue an active scene instead of starting again after a resume.

Calls

• Pass the selected env_id on every environment-scoped call: starting, checking and

finishing an episode, and every robot tool that robot-interface marks as scoped. Session-wide tools (status, the VLM, pure math) take none.

• In development Python, bind an environment once and call tools without repeating

env_id; keep the binding in the development wrapper, so the controller's main(ctx) stays the same everywhere (autonomous-control). Give each concurrent Python worker its own Context(): requests on one connection run one at a time.

• Harness calls that freeze or evaluate the controller act on the whole session; finish

outstanding control in every environment first.

Environment rules

• Each environment has its own scene, step counter and episode. Starting one does not

affect the others, and every start, including a failed one, uses an episode from the shared budget. An idle environment is capacity, not a free episode.

• Keep multi-call sequences (a plan and its execution, a preview and its replay) in the

environment that produced them, and pass an observation step only to the environment that produced it. Step ids count separately per environment, so a wrong one can still be accepted: track the env_id yourself.

• Calls within an environment run one at a time; different environments run

concurrently. This is not a lock around a whole script.

• Tie observations, pixels, plans and evidence paths to the env_id and episode, not

just a step number. Other environments have other layouts: they are not copies for same-state comparisons.

For delegated work, read multi-agent-exploration.