RoboCasa GR1 Tabletop evaluation protocol

TabletopBenchmarkAdapter, registered as robocasa_tabletop, owns native reset, control and scoring. robocasa_tabletop_contract.json binds sources, controller limits, encoder identities and the provisioned object inventory. Policies use the shared interface.

Tasks, variants and fast/full scope

The official set has 24 tasks: six container-closing tasks and eighteen PosttrainPnPNovel...SplitA tasks. Other registered pretraining and Eval...SplitB tasks are outside that set. Only these four task identities are implemented, through two explicit fast presets:

Scope	Native index	Task	Valid layouts	Styles
`pilot` (default compatibility name)	0	`PnPCupToDrawerClose`	[4]	0–11
`pilot`	19	`PosttrainPnPNovelFromTrayToCardboardboxSplitA`	[0]	0–11
`tabletop-next-pair-v1`	1	`PnPPotatoToMicrowaveClose`	[2]	0–11
`tabletop-next-pair-v1`	2	`PnPMilkToMicrowaveClose`	[2]	0–11

Native environment IDs are gr1_unified/<task>_GR1ArmsAndWaistFourierHands_Env. The campaign plan selects one pair. Fast mode evaluates each preselected task once, retaining failures/errors without automatic retry. Existing protocol IDs retain pilot-0-19 or next-pair-1-2 for compatibility.

Native-aligned full evaluation is not implemented. It requires the remaining official task identities and an explicitly validated native variant/reset/episode schedule. Four tasks sampled once, or retries of failed tasks, are not full mode. Layout/style variants do not create additional task identities.

Environment and reset

Pin RoboCasa GR1 revision 4840e671596f93ca03651524b9f72ffb1aadfeff and Robosuite revision 51cc01785bab80ffeed20da15e67d7dd4140e76a. The runtime has been pinned with Python 3.10.21. Set PYTHONHASHSEED=0 before Python starts; the adapter checks a hash sentinel.

Use seed 20260907, a fresh native constructor and one explicit seeded reset. Native constructor initialization is retained. The pinned Gym factory does not forward its layout/style arguments, so the adapter does not claim to force actual style. Preserve the full task VALID_LAYOUTS × 12 styles candidate set and record actual native selection privately.

Do not reset-search, success-screen, substitute assets or shrink the object library. Reset-time native success is an error before policy control, not a reason to resample. Setup and the bounded asset-snapshot limitation are described in environment/README.md.

Observation allowlist

One uint8 256 × 256 RGB image, ego_view_bg_crop_pad_res256_freq20.
Native annotation.human.coarse_action language, including unlocked_waist:.
Exactly 29 measured joint encoders in radians: left arm 7, right arm 7, left hand 6, right hand 6, waist 3.

The native image pipeline vertically flips a 1280 × 800 egoview, crops rows 310:770 and columns 110:1130, resizes to 720 × 480, pads to square, then resizes to 256 × 256. Copy that output without additional preprocessing. Do not use the old model-specific 224-pixel resize or 58D sine/cosine state transform.

Hand encoder order is pinky, ring, middle, index intermediate joints, thumb proximal pitch and thumb yaw. Exact identities are in the contract's state.parts. Observation metadata is empty. No extra cameras, depth, object or end-effector ground truth, world root pose, calibration, reward/success, reset metadata or simulator handles cross the policy boundary.

Absolute 29-dimensional joint action

One action contains absolute joint-position targets in radians at 20 Hz, using the native BASIC composite controller. It is not Cartesian motion, normalized [-1,1] output or joint velocity.

Slice	Order	Bounds in radians
Left arm [0:7]	Shoulder pitch/roll/yaw, elbow pitch, wrist yaw/roll/pitch	Low [-3,0,-3,-3,-3,-1.5,-1.5]; high [3,3,3,0,3,1.5,1.5]
Right arm [7:14]	Same joint order	Low [-3,-3,-3,-3,-3,-1.5,-1.5]; high [3,0,3,0,3,1.5,1.5]
Hands [14:20], [20:26]	Pinky, ring, middle, index, thumb pitch/yaw	Low all 0; high [1.57,1.57,1.57,1.57,1.22,1.74]
Waist [26:29]	Yaw/pitch/roll	Low [-1.05,-0.52,-0.7]; high [1.05,1.22,0.7]

Each hand expands six commands to eleven actuators using [0,0,1,1,2,2,3,3,4,4,5]; bounds intersect coupled joint and actuator ranges. The adapter rechecks compiled mappings, limits and ordinary encoder identities after reset. Invalid commands are rejected, not clipped, rescaled, passed through IK or replaced by heuristics. Evaluator-only measured hold generation is not a policy action service.

Budgets and scoring

For new service deployments, explicitly select budget_mode="horizon" and control_step_cap=1000. This independent-control development track has protocol gr1-tabletop-pilot-0-19/rgb256-qpos29/control-1000/v2 for the pilot selection (or gr1-tabletop-next-pair-1-2/... for the next pair). A positive integer cap at most 1000 is allowed; each cap has its own control-N/v2 identity. The service publishes max_policy_calls=null and replan_interval=null: the policy/method chooses chunk length and the trusted driver may declare its own call or replan limit. Each service step still executes one native control; native success/termination and the selected total cap stop execution even in the middle of a chunk. Service decision-limit parameters cannot be combined with this track.

The omitted/default budget_mode="decisions" preserves the historical 200-call, prefix-5, D × K control budget and its existing decision-D-prefix-K/v1 identity. Its D=1..200 and K=1..5 limits, smaller budgets and frozen campaign plans remain unchanged. control_step_cap is not accepted in that track. These two budget identities must not be pooled as one protocol.

Both remain provisional development tracks. The native engine stays at horizon 1000 with ignore_done=True; neither re-labels that engine setting as a verified native-full evaluator budget or establishes full-native protocol parity.

Each adapter step executes one native control. Native reward uses _check_success(), and Gym returns info.success = reward > 0. Consume the flag without a second oracle query and stop at first native success. Preserve native termination and truncation separately.

The supported scorers are stateless current-scene predicates. CupToDrawer checks containment and drawer closure. Novel placement checks gripper separation, receptacle placement, counter contact and uprightness. The microwave tasks require the object's center inside the microwave (partial_check=True) and normalized door fraction at most 0.005. These are evaluator-private scoring definitions, not policy feedback or alternate visual judges.

Record actual controls, policy calls, action decisions and first-success control/call index. A stop-only response has no action decision; shorter/truncated chunks invalidate division-by-five estimates. A minimum-actions-before-voluntary-stop preference belongs to the model profile and cannot delay native success, evaluator caps or errors.

Evidence and acceptance limits

Retain source/model/protocol identities, actual task/layout/reset and native metadata, compiled scene XML, initial simulator state/hash, controller/encoder manifest, used-asset closure, delivered frames, action chunks/executed prefixes, per-control metrics, errors and decodable videos. Keep private native artifacts out of the policy mailbox. Store model outputs under AgentPolicies/<profile>/results/robocasa_tabletop/.

The object contract certifies a provisioned snapshot, not the complete official asset release. Preserve native missing-category warnings and catch-and-skip behavior, including possible sketchfab/book distractor warnings; validated used-scene assets do not erase that limitation. Do not silently modify the library to clear a warning.

Source checks, controller parity, sensor allowlisting, per-action scoring and video decoding are acceptance requirements. Zero-policy reset/hold diagnostics and CPU fixtures are infrastructure evidence, never capability attempts or complete 24-task coverage.