common/source/starBenchserver/benchmarks/simBenchmarks/MetaWorld/README.md

MetaWorld MT50 SIM server

The benchmark owns the native simulator, RGB/action contract and success check. Models own only their observation-to-action policy under AgentPolicies/<profile>/; the neutral harness owns communication, counters and reports.

This is a development server adapter, not a completed native MT50 score. The current protocol supports an explicit fast subset, one attempt per selected task. Native full evaluation and retained-evaluator parity are not established.

• Environment setup and package verification
• Evaluation protocol
• All 50 canonical task names, instructions and retained difficulty buckets
• Shared SIM service interface
• Example fixed server configuration

Directory responsibilities

server.py is the public factory/launch entrypoint. bench_server/adapter.py owns observation, action and native scoring; bench_server/protocol/task_catalog.json owns the immutable task inventory. environment/ owns native setup and package pins; checks/ contains offline contract regressions. Deployment paths and the selected schedule belong in an operator copy of server.example.json. There is no root adapter compatibility entrypoint. This layout changes neither the public wire protocol nor native control/scoring semantics.

Launch a fixed schedule

Use the prepared native Python, from a stable source checkout. Set a strong RSI_SIM_SERVER_TOKEN outside Git in the server and trusted evaluator only. Never give the token, service management interface or native filesystem to a model actor. Choose a fresh result directory for each authorized schedule.

python -B -m benchmarks.simBenchmarks.MetaWorld.server \
  --parameters benchmarks/simBenchmarks/MetaWorld/server.example.json \
  --seed 4042 --episodes 1 --port 8765 \
  --output-dir /absolute/AgentPolicies/my_profile/results/metaworld/raw/reach-direct-control/server

Launch/list/describe do not construct the native simulator. The first claimed episode's reset constructs MT50, selects the task and performs native setup. The service binds loopback only; use SSH forwarding for a separate evaluator node. There is no background control loop and no reset/action replay on errors.

From the evaluator environment, with the policy already configured:

python -B -m benchkit run --benchmark metaworld \
  --sim-server http://127.0.0.1:8765 \
  --model-profile anonymous_code_agent \
  --policy-args '{"directory":"/absolute/opaque-mailbox"}' \
  --seed 4042 --episodes 1 --budget-mode horizon \
  --max-steps 400 --max-policy-calls 200 --replan-interval 5 \
  --output-dir AgentPolicies/anonymous_code_agent/results/metaworld/raw/reach-direct-control/evaluator

Use horizon here: the cap is 400 executed model-selected controls, not a 200×5=1000-control extension. A 200-call cap or shorter policy chunks can end an attempt earlier. Stop calls execute no control. The one native setup no-op is recorded separately. Ordinary give-up preferences belong to the policy, not this benchmark.

The actor receives full-frame 480×480 corner2 RGB and the instruction, never native object/goal state. The default image_orientation="rot180" rotates the native upside-down view exactly 180 degrees; it performs no crop or resize. The explicit native orientation preserves the earlier raw fast-v1 track under its original protocol ID. These are distinct observation conditions, recorded in provenance, not interchangeable scores. The actor selects a four-dimensional bounded native action. No model server or VLA-specific preprocessing is used. Old model scripts remain in Git history; the adapter never imports them.

All per-run results, evidence and videos belong to the selected profile's results/metaworld/, not this benchmark directory. A failed fast task is retained; changing directories or restarting a service does not authorize another attempt.