common/source/starBenchserver/benchmarks/simBenchmarks/Robotwin/EVALUATION_PROTOCOL.md

RoboTwin clean direct-control protocol

Protocol: robotwin-local-content/clean-fixed-seed/qpos14-rgb320-state14/fast-v3. Opt-in renderer compatibility protocol: robotwin-local-content/clean-fixed-seed/qpos14-rgb320-state14-oidn233/fast-v4. Status: development server adapter; native acceptance requires separate evidence. Benchmark registration name: robotwin.

These versions add a declared robot-state channel to the former RGB-only fast-v1 / fast-v2 tracks. Old frozen results retain their original inputs and identities; this is not a silent reinterpretation or a claim of score parity. Native cameras, actions, control timing, task selection and scoring are unchanged.

Scope and denominator

The retained launcher lists 50 tasks and clean/randomized settings. The inspected native script/eval_policy.py defaults to 100 accepted test episodes per task, not the episode_num: 50 collection field in demo_clean. It starts at 100000 * (1 + seed), performs expert play_once feasibility screening, skips unsuccessful seeds, then resets accepted seeds again for the participant. Language templates may use expert-generated episode information.

This server does not execute that expert/filtering loop. Its explicit fast development track chooses one task and one unscreened native seed in advance, then calls setup_demo once. Failure, unstable-scene errors and runtime errors are retained; none trigger another seed or attempt. Native object-generation sampling within that one seeded scene setup is unchanged. Language is the task JSON's fixed full_description, without expert execution, hidden object identifiers or adaptive prompt selection.

Implemented task	Native object variants retained	Native target-control horizon
lift_pot	060_kitchenpot, IDs 0 and 1; native seeded sampling	400
click_bell	050_bell, IDs 0 and 1; native seeded sampling	400

Only demo_clean with native Aloha is exposed. A service has one selected task, one declared native_seed and one attempt; a second open/reset is rejected even after failure. Full mode, EEF actions, other tasks and arbitrary native config overrides are unsupported. This denominator is not native-full coverage and is not directly comparable with expert-filtered scores. A future full track needs a reviewed native schedule, not success-seeking retries during evaluation.

Source, runtime and configuration identity

The native local archive has no own Git metadata. Its selected files are pinned in environment/native_snapshot.json, inventory SHA-256 ab5a73b0482d50d9c83e8702b37e7e7eb10bf7033e03a8fec59b60b708bef08c. native_source_revision is null; the descriptor's required identity string is content-inventory-sha256:<digest>. A parent repository's revision or current upstream HEAD is not a substitute for this archive.

The deployment helper copies matching source/assets to a new owned root. It permits exactly native Aloha left/right CuRobo template path expansion; original config/template hashes, derived bytes, root and manifest hash are recorded. Other bytes and the file set remain exact. Runtime metadata and the selected SAPIEN/MPLib/TOPPRA package file inventory are verified before native imports. Drivers, JIT outputs, all transitive dependencies and loaded memory are not attested by these checks.

The default renderer_runtime: bundled_oidn_2_0_1 preserves the original runtime snapshot and protocol. The explicit oidn_2_3_3_cuda_v1 selection uses an owned SAPIEN 3.0.0b1 overlay with the official OIDN 2.3.3 CUDA libraries. Only the OIDN library directory and two literal paths in SAPIEN's OIDN loader change; native RoboTwin code/assets, SAPIEN physics/renderer binaries, shaders, 32 samples per pixel, depth 8, camera definitions and scorer remain unchanged. It does not disable denoising, select OptiX/CPU, or install task-solving code.

The compatibility inventory additionally checks all three sapien.libs binaries, preserved distribution metadata and regenerated RECORD, rejects unlisted overlay files and preloaded native libraries, and checks actual loaded library paths after import and the first observation. Other dependencies remain read-only references to the original selected environment; this is not a fully isolated or transitively pinned environment. Verification receipts are saved before native import so an initialization failure retains its runtime identity. See renderer setup for preparation and acceptance.

OIDN 2.3.3 supports newer CUDA architectures but has different denoiser implementation/default filter quality from 2.0.1. The compatibility protocol therefore has a distinct identity: no pixel parity or interchangeable scores between these tracks are claimed. A device/filter check is not native rendering validation; separately inspect the declared server diagnostic and all cameras.

Observation boundary

The actor receives language plus head_camera, left_camera, right_camera in that order, each native 320 × 240 RGB uint8. There is no flip, resize, crop, overlay, image annotation or channel conversion. The native camera already converts its Color buffer to uint8 RGB.

At reset and after every acknowledged control, state contains 14 robot values in the following fixed order, declared individually in state_names:

Indices	Fields	Source and units
0–5	measured_fl_joint1_rad … measured_fl_joint6_rad	Actual left-arm articulation joint positions, radians
6	commanded_left_gripper_01	Last native normalized left-gripper command; 0 closed, 1 open
7–12	measured_fr_joint1_rad … measured_fr_joint6_rad	Actual right-arm articulation joint positions, radians
13	commanded_right_gripper_01	Last native normalized right-gripper command; 0 closed, 1 open

The pinned native robot's get_left_arm_real_jointState and get_right_arm_real_jointState read arm positions through articulation get_qpos in the compiled arm-joint order. Each getter appends its native gripper command cache. Thus there are 12 measured arm values and two commanded gripper values, not 14 measured encoders. Gripper commands do not establish physical opening, contact or a successful grasp. These read-only getters do not advance physics or invoke task logic; the state and cameras are sampled at the same paused control boundary. Measured arm values are never clipped to action bounds. Missing, non-finite or wrongly shaped state is an explicit error, not a fabricated zero vector or a fallback to commanded joints.

The similarly named native joint_action.vector still contains drive targets; it is not used as measured state. Only the declared camera arrays and robot state are projected. Extra front camera, depth, point clouds, segmentation, calibration, EEF/object poses, seeds, task metadata, native score and debug channels remain excluded. No arbitrary native observation dictionary is forwarded.

Direct action boundary

The actor selects this 14-D vector directly:

[fl_joint1, fl_joint2, fl_joint3, fl_joint4, fl_joint5, fl_joint6, left_gripper,
 fr_joint1, fr_joint2, fr_joint3, fr_joint4, fr_joint5, fr_joint6, right_gripper]

Arm entries are absolute native joint radians with authored URDF bounds [-10, 10]. Grippers are absolute normalized commands in [0, 1] (0 closed, 1 open), using the native [-0.01, 0.045] meter scale. At reset the adapter checks actual compiled joint names/limits, gripper scale, horizon and physics timestep. A mismatch is a contract/environment error before actor controls, not permission to infer different bounds.

One take_action(vector, action_type="qpos") executes each control. Native MPLib TOPP interpolates from previous drive targets to requested absolute joint targets under fixed low-level velocity/acceleration constraints. It does not select a grasp or plan around task objects. EEF/CuRobo planning, expert methods, VLA proposals, fallback task policies and task-solving services are not exposed.

Native setup still constructs planners, moves to homestate, opens grippers and performs object-stability settling. These are declared initialization, not model decisions. No play_once, expert feasibility check, hidden-state action generation or scored-seed search occurs.

Native qpos control catches TOPP exceptions/zero-sample paths and uses its fixed 50-tick gripper interpolation for the affected arm. This existing low-level behavior is preserved and explicitly recorded by passive TOPP instrumentation; the adapter never substitutes a task action or replays a failed control. Exceptions outside that native controller retain errors.

Timing, stopping and success

Physics uses native 1/250-second ticks. A target consumes a variable number according to TOPP; control_hz is unspecified, not inferred from video FPS. Native take_action_cnt stops at 400 for both tasks, from the pinned _eval_step_limit.yml. Runner decision limits do not extend that horizon. The example uses 200 policy calls, an action prefix of at most 5 and a separate 1,000-control ceiling.

Native take_action checks check_success during each physics tick and caches eval_success. The adapter consumes that result without adding a second stateful scorer query after each target or inferring success from video. Numeric transition reward only encodes this Boolean success; no shaped reward is introduced. Success terminates immediately. Native horizon, runner limits, errors and explicit policy stop can end unsuccessfully. Voluntary persistence and minimum-decision rules remain policy-owned.

Record calls, executed action decisions, native started target controls, acknowledged transitions and completed physics ticks separately. A passive task-side scene view counts physics after reset. Initialization ticks are unknown, not falsely zero; setup includes settling and fixed controls. On a native exception, partially executed controls may have consumed physics without an acknowledgment. Preserve that uncertainty and never replay the action.

Media, audit and lifecycle

Native events, resolved config/provenance, delivered-frame PNG/RGB hashes, three named-camera MP4s, counters and cleanup receipts remain evaluator-private under AgentPolicies/<profile>/results/robotwin/. The actor receives only public observations and execution feedback through its policy interface. Idle wall time does not advance simulation.

MP4s present one public frame per control at 10 FPS. They are neither real-time physics captures nor proof of motion between samples. frames.jsonl records completed physics ticks/simulated seconds beside each frame. Include camera identity and task descriptions in exported paths; distinguish reset-only diagnostics from model attempts.

Close calls native close_env and releases references, retaining errors. Native gym close does not prove GPU resource reclamation. Use one episode per server process and stop it after result transfer; do not reuse it for a task. Logical policy/evaluator boundaries are not an OS sandbox.

Acceptance gates requiring native evidence

CPU tests validate factory/lifecycle/projection/deployment and passive counters, not SAPIEN rendering, physics or model performance. Before scoring, independently review and run a separate one-reset/two-fixed-control infrastructure fixture; verify compiled limits and delivered cameras, decode videos and audit counts and cleanup. Keep the fixture out of model success rates. Randomized/full evaluation and serial/sharded native parity need further implementation and benchmark-specific evidence.

Additive official-source deployment track

The explicit deployment_track: official_ea8b211_sapien303_v1 selects robotwin-official-ea8b211/clean-fixed-seed/qpos14-rgb320-state14-sapien303/fast-v1. It does not require the historical local archive. The default local_archive track and its original snapshots remain unchanged.

This track pins official RoboTwin commit ea8b21121ebb3cd201ff5b3fe361944ac94eda3f; the selected source closure is listed in environment/official_source_snapshot.json. The official evaluator is scripts/eval_policy_xpolicylab.py; configurations live in env_cfg/task_config. Its task classes, measured-state getters, qpos controller, native TOPP behavior, 400-target limits and native success predicates are used without source edits. The service still selects one unscreened native seed and never invokes the expert screening/policy-server path of that evaluator. This is a development track, not an official full-suite score or parity claim with the archived track.

Official assets come from the upstream downloader's Hugging Face dataset TianxingChen/RoboTwin2.0, pinned to revision 3dc3b798668feb99ac61cc9086d84cbcc3d79186. The asset receipt binds the downloaded embodiments.zip and objects.zip SHA-256 values plus extracted clean-task files. Only Aloha, bell/pot and required object metadata are copied. Background textures and unrelated task assets are unnecessary for the declared clean mode. The 1,547 selected extracted files have canonical inventory SHA-256 3e0bb720efcf4a516cbaa14af27f50803d33c2ac4124a0bf25f2ebdbde95c3f5, fixed in the implementation; a receipt cannot substitute arbitrary file hashes while merely retaining the two real archive hashes. Only the two native CuRobo template path expansions are derived during deployment.

The runtime is explicitly Python 3.10.12, SAPIEN 3.0.3, MPLib 0.2.1, TOPPRA 0.6.3, NumPy 1.26.4 and Torch 2.5.1+cu121. These differ from upstream's installation recommendations and therefore have a distinct protocol identity. Preparation binds captured SAPIEN/MPLib/TOPPRA/CuRobo package bytes, including sapien.libs when present; readiness verifies those exact bytes and versions again before native imports. CuRobo remains required by upstream Aloha initialization even though actor control uses qpos/MPLib TOPP. No planner fallback is introduced. The receipt does not attest the GPU driver, every transitive dependency, JIT outputs or loaded memory. Native acceptance remains the separate one-reset, two-fixed-control diagnostic, including camera, counter, video and cleanup checks.