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

RoboCasa GR1 Tabletop

Directory responsibilities

Robocasa_tabletop/
├── server.py, server.example.json   # public launcher and operator configuration
├── bench_server/
│   ├── adapter.py                  # public observations/actions and native scorer
│   └── protocol/
│       ├── plan.py                 # predeclared fast selection
│       ├── robocasa_tabletop_contract.json # pinned sensor/action/source contract
│       └── POLICY_GUIDE.md           # public robot interface, not a model prompt
├── environment/                    # native readiness, runtime pins and helpers
└── checks/                         # dependency-light conformance regressions

Protocol catalogs are benchmark definitions, not datasets or deployment settings. Source/asset locations and GPU selection remain operator configuration. There is no root adapter alias or separate evaluation/ package; current callers use bench_server.adapter and bench_server.protocol.plan. The public server module name is unchanged. This directory migration changes no native controls, observations, scoring, budgets, selected cases or historical evidence. Model behavior and results belong under AgentPolicies/<agent>/; shared communication belongs in benchkit/sim_server/. Structure and offline checks do not establish native service readiness or native-full evaluation support.

The robocasa_tabletop adapter evaluates GR1 whole-upper-body manipulation using native RoboCasa GR1 Tabletop tasks and success predicates. This is distinct from PandaOmron RoboCasa365.

The official set contains 24 tasks. The implemented fast selections are two explicit pairs: indices 0/19 or 1/2. Other task identities and native-aligned full evaluation are not implemented. See EVALUATION_PROTOCOL.md for task names, variants, sensors, physical joint actions and scoring.

Prepare and verify

Follow environment/README.md for source/runtime/asset requirements and a parameterized native diagnostic. Required source pins are:

• RoboCasa GR1: 4840e671596f93ca03651524b9f72ffb1aadfeff.
• Robosuite: 51cc01785bab80ffeed20da15e67d7dd4140e76a.

Launch Python with PYTHONHASHSEED=0 before process startup. The pinned contract includes the complete provisioned object inventory and compiled controller/encoder identities; arbitrary newer source or asset snapshots are not interchangeable. A portable all-assets bootstrap is not yet available.

CPU checks do not launch a native simulator:

python -B -m unittest discover -s benchmarks/simBenchmarks/Robocasa_tabletop/checks -v

Run as an independent SIM server

server.py owns the native Tabletop adapter. The evaluator connects through the shared SIM client; the policy receives only the one allowed RGB view, native language and 29 measured encoders. No server handle, private metadata, scorer or diagnostic hold is a policy interface.

Copy server.example.json to an operator-owned file and replace all three source/inventory placeholders and the GPU before launch. These paths must still satisfy the exact source/asset contract; the example does not provision or certify a fresh installation. It selects task 0 PnPCupToDrawerClose, one seeded reset and the explicit budget_mode: "horizon", control_step_cap: 1000 development track. This server track limits total controls, not policy calls or controls per decision. The policy/method owns its chunk size and any replan/call limits. It is not native-full evaluation.

Follow SERVER_PROTOCOL.md to set the same strong RSI_SIM_SERVER_TOKEN only in server/evaluator process environments and to forward loopback over SSH when using separate nodes. Start the server from a stable source snapshot using the pinned native Python 3.10.21 environment:

PYTHONHASHSEED=0 python -B -m benchmarks.simBenchmarks.Robocasa_tabletop.server \
  --parameters /absolute/path/to/tabletop-server.json \
  --seed 20260907 --episodes 1 --port 8768 \
  --output-dir AgentPolicies/anonymous_code_agent/results/robocasa_tabletop/raw/PnPCupToDrawerClose-cup-in-drawer/server

PYTHONHASHSEED=0 must be present before the server interpreter starts; changing it inside Python cannot satisfy the native hash sentinel. The chosen GPU must be idle at reset. Do not stop another user's process to pass the GPU check. In a separate evaluator environment with NumPy and an unused opaque mailbox actor endpoint:

python -B -m benchkit run \
  --benchmark robocasa_tabletop --sim-server http://127.0.0.1:8768 \
  --model-profile anonymous_code_agent \
  --policy-args '{"directory":"/absolute/opaque-mailbox","timeout_s":900,"max_chunk_length":5}' \
  --seed 20260907 --episodes 1 --budget-mode horizon \
  --max-policy-calls 200 --max-steps 1000 \
  --output-dir AgentPolicies/anonymous_code_agent/results/robocasa_tabletop/raw/PnPCupToDrawerClose-cup-in-drawer/evaluator

In this example the mailbox method chooses its own five-action maximum and 200-call allowance; neither is a server constraint. Omitting the generic runner's --replan-interval executes the submitted chunk until native end or the remaining total control limit. A method may instead declare its own prefix. The service accepts control_step_cap integers from 1 to 1000; a smaller cap has a distinct control-N/v2 protocol identity. Do not mix that parameter with the legacy service-side max_policy_calls or replan_interval parameters.

Do not pass --benchmark-args remotely; only the service constructs native environments. Use new output directories. Service launch/describe freezes the selection without resetting the SIM and is not native readiness evidence. Source, inventory, renderer, compiled controller and encoder checks remain in the adapter's one reset. Failure retains the attempt; no automatic retry or asset substitution is enabled. Keep original service/client records and videos under the model result namespace and verify source integrity before and after. The evaluator does not need the native Python version or renderer settings. Native-free service checks do not establish native service-mode validation, full 24-task support or object-library completeness. The frozen campaign path and legacy artifact auditors do not yet support remote service records.

In-process compatibility

The adapter default and existing frozen campaigns retain the separately named decisions track, including its 200-call/prefix-5 defaults and D × K total cap. This legacy example is not the recommended independent-control server track. After provisioning compatible sources/assets and connecting a mailbox actor:

PYTHONHASHSEED=0 python -m benchkit run \
  --benchmark robocasa_tabletop --model-profile anonymous_code_agent \
  --benchmark-args '{"task_ids":[0],"casa_source":"/path/to/robocasa-gr1","suite_source":"/path/to/robosuite-gr1","object_inventory":"/path/to/object-inventory.json","render_gpu_device_id":0,"budget_mode":"decisions","max_policy_calls":200,"replan_interval":5}' \
  --policy-args '{"directory":"/path/to/opaque-mailbox"}' \
  --seed 20260907 --episodes 1 --budget-mode decisions \
  --max-policy-calls 200 --replan-interval 5 --max-steps 1000

Explicit paths avoid the adapter's site-specific defaults; the files must still match the pinned manifest. Use task_scope="tabletop-next-pair-v1" with task_ids=[1] or [2] for the second pair. The campaign plan supports either pair and requires compatible deployment paths.

Fast mode keeps one outcome for every preselected task, including failure and environment error, without automatic retry. Only the evaluator operates the simulator and reads native scoring state. Model behavior/results belong under AgentPolicies; anonymous transport is not proof of a named model/provider identity.