Benches

A bench is a simulator worker: one task per process, served over the worker protocol. Isaac, torch, cuRobo and the simulator are imported only in the worker process.

RoboDojo

The dual-arm ARX X5 of RoboDojo, at the revision in third_party/sources.toml.

benches/robodojo/
├── adapter.py     RoboDojoEnvironment: native reset and take_action, cameras, action packing, reward
├── general.py     RoboDojoTaskEnvironment: any one task, evaluation or development layouts
├── runtime.py     RoboDojoRuntimeEnvironment: the robot API, the shield
├── kinematics.py  cuRobo FK, IK and plans in the env frame (also the kinematics service)
├── devlayouts.py  development layouts, disjoint from the evaluation packs
└── codec.py, tasks.py, video.py

harness serve --factory benches.robodojo.runtime:create_environment \
  --options configs/workers/robodojo_runtime_dev.toml --port 8771

harness run starts and stops workers itself. By hand, set HARNESS_ROBODOJO_RUNTIME=1, an absolute HARNESS_ROBODOJO_ROOT (the worker changes directory before it imports RoboDojo), CUDA_VISIBLE_DEVICES, a writable TMPDIR (Isaac Lab logs under $TMPDIR/isaaclab) and OMNI_KIT_ACCEPT_EULA=YES (source environment/activate.sh sets the last). Isaac ignores SIGTERM: stop a worker with SIGINT, then SIGKILL.

Action and observation

• A command is 14 values: components 0-5 and 7-12 are absolute joint1-joint6 targets in radians (left arm, then right), 6 and 13 normalized gripper openings in [0, 1] (1 open; the tasks count at least 0.8 as open).
• One tick is one policy step at 25 Hz; RoboDojo interpolates it over its control steps.
• Reward follows RoboDojo’s episode score: 1 when the success flag passes, otherwise the process score / 100. Reaching the step limit without success is truncation.
• Cameras cam_head, cam_left_wrist, cam_right_wrist, 640×480 RGB; no depth (the evaluation configuration renders none). End-effector poses are link6 poses.
• An unstable layout, or an episode a scripted support arm marks unstable, is an environment error: neither a success nor a failure of the policy.

Layouts

• Evaluation: layout_source = "eval", layout_pack_seed 0/1/2, the official packs.
• Development: layout_source = "procedural" with dev_layout_dir (default <checkout>/harness_dev_layouts/arx_x5) replays <dir>/<task>_<seed>.json for any seed, generating it on first use (scripts/robodojo_make_dev_layouts.py pre-generates them without Isaac Sim). The public RoboDojo has no layout generator, so devlayouts.py re-places a template evaluation layout with RoboDojo’s own ClutteredGenerator within each instance’s config limits, redraws object choices and clutter, and rejects any layout that repeats an evaluation layout. The same (task, seed) gives a byte-identical file. On the two conveyor tasks the objects on the belt are re-placed in its slots and turned a little; objects off it keep their pose. Not supported: cover_blocks, play_tic_tac_toe, swap_blocks, whose configs leave no placement freedom beyond the evaluation packs.

Scene configurations follow RoboDojo’s task/RoboDojo/config/_task.yml (conveyor for the conveyor tasks, a second scripted robot for imitate_sorting_sequence, make_kong, play_tic_tac_toe); the policy always drives the two X5 arms. As RoboDojo’s run_eval does, the adapter loads the scripted arm’s trajectory after reset, before the policy’s first step (the opponent’s first move, the demonstration). The PhysX crash monitor stays off: a crash ends the worker, the driver records the error, and harness run --rerun-errors runs the episode again.

Measured facts the prompts use

• Control at 25 Hz; a tick takes about 0.14 s of wall time (0.19 s with the head camera at 5 Hz in a controller, 0.4 s at 25 Hz). The simulator waits while programs think.
• link6 reaches forward only to y ≈ +0.1; pointing straight down ([0.5, -0.5, 0.5, 0.5]) needs z ≥ 0.9 and y ≤ 0.0 (scripts/robodojo_reach_probe.py).
• Table top z = 0.765. Fingertips are the centre of the gripping faces, 0.150 m along link6 +x (the jaws end 0.0075 m further: pointing down they touch the table at link6 z ≈ 0.923); fingers slide along link6 y, 0.044 m each; an empty closed gripper reads an opening of about 0.185.
• Motions are timed at the planner’s joint velocity and acceleration limits (3.14 rad/s, 10 rad/s²); a 15 cm move takes about 33 ticks at the planner’s timing, 22 at speed=0.3.
• The camera calibration is the same for every task, layout and arm pose (head camera fixed in the world, wrist cameras fixed on link6): embodiments/arx_x5.

RoboTwin 2.0

benches/robotwin/runtime.py: the same robot API at 10 Hz, 50 tasks, seeds checked by RoboTwin’s expert, its own cuRobo 0.7.8. Run on RoboTwin 2.0.

harness serve --factory benches.robotwin.runtime:create_environment \
  --options configs/workers/robotwin_runtime_dev.toml --port 8772