{ "schema_version": 1, "status": "source audit; not experimental outcomes", "source_revision": "81cf7909b13fe1a961aae9a4e457a5e786e4df8c", "arena_revision": "c5826c87aa181442328c2c74c49f9e551b5c3d8c", "expert_revision": "e0f5a056257fc89b386b7584cd5978285c6a02de", "dataset_revision": "1c5b2fbfaa35f70451bb462426c1b8adec1b7ec6", "assets_revision": "e005c6be4e8b5522c7efe1040764b97df7fcdc38", "task_catalog": [ { "id": "LMBoxLift", "instruction": "Walk to the table and pick up the box", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Object root world Z >0.18m and at least one finger link within0.2m.", "root_z_gt_m": 0.18, "finger_near_m": 0.2 }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "table_center_xyz_m": [ 2, 0, -0.8 ], "table_xy_half_range_m": 0.16, "table_yaw_center_half_range_rad": [ 3.141592653589793, 0.16 ], "object_offset_from_table_world_axes_m": { "x": [ -0.41, -0.09 ], "y": [ -0.16, 0.16 ], "z": [ 0.80145, 0.80145 ] }, "object_yaw_center_half_range_rad": [ 1.5707963267948966, 0.16 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMBoxLift.py", "source_sha256": "1f73c5e0af8f159aeffc405495032db4f658c753bbecc328ddbf12a847f64261", "interpretation_caveat": "Absolute world root height plus finger proximity; no secure-contact or dwell requirement." }, { "id": "LMBoxLiftFloor", "instruction": "Walk to the box and pick it up from the floor", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Object root world Z >\u22120.5m, local-Z/world-up dot \u22650.9, and at least one finger within0.6m.", "root_z_gt_m": -0.5, "upright_dot_ge": 0.9, "finger_near_m": 0.6 }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "box_center_xyz_m": [ 2, 0, -0.8 ], "box_xy_half_range_m": 0.16, "box_yaw_center_half_range_rad": [ 0, 0.16 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMBoxLiftFloor.py", "source_sha256": "8c1625ef2310437afbe343a2c2a29b20bf5adb360e41b519603758411ce84e35", "interpretation_caveat": "Absolute root height and orientation; the0.6m finger proximity allowance is a broad proxy." }, { "id": "LMBoxTableToShelfPnP", "instruction": "Move the box from the table to the shelf", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Object/destination root separation per axis <[0.16,0.14,0.9227]m, object static, every finger >0.1m away, and destination contact >0.95N.", "root_separation_lt_xyz_m": [ 0.16, 0.14, 0.9227 ], "static": true, "all_fingers_far_m": 0.1, "destination_contact_gt_N": 0.95, "drop_below_root_z_m": -0.6 }, "native_reset": { "robot_fixed": true, "source_table_center_xyz_m": [ 1.4, 0, -0.8 ], "source_table_xy_half_range_m": 0.1, "source_table_yaw_center_half_range_rad": [ 3.141592653589793, 0.02 ], "object_offset_from_table_world_axes_m": { "x": [ -0.341, 0.159 ], "y": [ -0.278, 0.322 ], "z": [ 0.762, 0.762 ] }, "object_yaw_center_half_range_rad": [ 0, 0.36 ], "shelf_center_xyz_m": [ 0.3, -1.1, -0.8 ], "shelf_xy_half_range_m": 0.18, "shelf_yaw_center_half_range_rad": [ 1.5707963267948966, 0.36 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMBoxTableToShelfPnP.py", "source_sha256": "c4a1540f34cb3b8ce4e4d0c7fe5f8e1fb80f12eef1728236ccb9d697d7d2c9b4", "interpretation_caveat": "World-axis root proximity and destination contact; static permits speed up to0.15m/s and0.15rad/s." }, { "id": "LMDrillLift", "instruction": "Walk to the table and pick up the power drill", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Drill root world Z >0.16m and approximate grasp predicate.", "root_z_gt_m": 0.16, "approximate_grasp": true }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "table_center_xyz_m": [ 2, 0, -0.8 ], "table_xy_half_range_m": 0.16, "table_yaw_center_half_range_rad": [ 3.141592653589793, 0.1 ], "drill_offset_from_table_world_axes_m": { "x": [ -0.38, -0.1 ], "y": [ -0.4, 0.4 ], "z": [ 0.80145, 0.80145 ] }, "drill_yaw_center_half_range_rad": [ -1.5707963267948966, 0.4 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMDrillLift.py", "source_sha256": "6a0f3db2ca052b480fb79167e7de3226023fd22def58336627dacf21567aca94", "interpretation_caveat": "Grasp is approximated by finger proximity and joint closure, without required object-hand contact." }, { "id": "LMDrillLiftObs", "instruction": "Navigate around the shelf and pick up the power drill from the table", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Drill root world Z >0.16m and approximate grasp predicate. Success does not explicitly require obstacle avoidance.", "root_z_gt_m": 0.16, "approximate_grasp": true }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "obstacle_center_xyz_m": [ 0.805, 0, -0.8 ], "obstacle_xy_half_ranges_m": [ 0.12, 0.6 ], "obstacle_yaw_half_range_rad": 0.12, "table_center_xyz_m": [ 2.3, 0, -0.8 ], "table_xy_half_range_m": 0.12, "table_yaw_center_half_range_rad": [ 3.141592653589793, 0.12 ], "drill_offset_from_table_world_axes_m": { "x": [ -0.35, -0.14 ], "y": [ 0.28, 0.64 ], "z": [ 0.80145, 0.80145 ] }, "drill_yaw_center_half_range_rad": [ -1.5707963267948966, 0.3 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMDrillLiftObs.py", "source_sha256": "c8d4ba247813dab5d62370d574ef26e5702c6ae90ddf5ece3c4c84490f56d278", "interpretation_caveat": "Approximate grasp; success does not explicitly test that the robot avoided colliding with the obstacle." }, { "id": "LMDrillPnP", "instruction": "Move the power drill from one table to another", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Drill/destination-table root separation <[0.3,0.27,0.81]m, static, local-Z/world-up dot \u22650.8, and every finger >0.1m away. No contact term.", "root_separation_lt_xyz_m": [ 0.3, 0.27, 0.81 ], "static": true, "upright_dot_ge": 0.8, "all_fingers_far_m": 0.1, "destination_contact_required": false, "drop_below_root_z_m": -0.6 }, "native_reset": { "robot_fixed": true, "source_table_center_xyz_m": [ 1.2, 0, -0.8 ], "source_table_xy_half_range_m": 0.12, "source_table_yaw_center_half_range_rad": [ 3.141592653589793, 0.12 ], "drill_offset_from_table_world_axes_m": { "x": [ -0.3, -0.14 ], "y": [ -0.3, 0.3 ], "z": [ 0.8, 0.8 ] }, "drill_yaw_center_half_range_rad": [ -1.5707963267948966, 0.3 ], "destination_table_center_xyz_m": [ 0, 1.2, -0.8 ], "destination_table_xy_half_range_m": 0.12, "destination_table_yaw_center_half_range_rad": [ -1.5707963267948966, 0.12 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMDrillPnP.py", "source_sha256": "b482efbf037b23173ee36edbbb64de370e0c91b7a539a91faf7f60326c3d4063", "interpretation_caveat": "Root proximity, orientation and static test; no destination contact term." }, { "id": "LMPickDrillFromHolder", "instruction": "Pick up the power drill from the holder", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Drill root world Z >0.13m, approximate grasp, and holder contact is not above0.95N.", "root_z_gt_m": 0.13, "approximate_grasp": true, "holder_contact_not_gt_N": 0.95 }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "shelf_center_xyz_m": [ 2, 0, -0.79 ], "shelf_xy_half_range_m": 0.12, "shelf_yaw_fixed_rad": 1.5707963267948966, "holder_offset_world_axes_m": { "x": [ -0.769, -0.769 ], "y": [ -0.12, 0.12 ], "z": [ 0.75, 0.75 ] }, "drill_slot_world_offsets_m": [ [ -0.07, 0.165, -0.16 ], [ -0.07, 0.055, -0.16 ], [ -0.07, -0.055, -0.16 ], [ -0.07, -0.165, -0.16 ] ], "slot_rng": "Python random.choice; also seed NumPy and Torch for other native reset samplers" }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMPickDrillFromHolder.py", "source_sha256": "08ef91d0eee6ad68203cb6d1b81209f993af8bd4b4619f673b6118ba89f4d3f5", "interpretation_caveat": "Approximate hand grasp plus no holder contact; removing the drill is not a secure-grasp guarantee." }, { "id": "LMPushButton", "instruction": "Walk to the console and press the red button", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Button joint normalized pressedness >0.5 and at least one finger within0.2m.", "normalized_pressedness_gt": 0.5, "finger_near_m": 0.2, "reset_normalized_pressedness": 0.2 }, "native_reset": { "robot_xy_half_range_m": 0.05, "robot_yaw_half_range_rad": 0.17453292519943295, "console_world_ranges_m": { "x": [ 1, 1.8 ], "y": [ -0.2, 0.8 ], "z": [ -0.79, -0.79 ] }, "console_yaw_fixed_rad": 1.5707963267948966, "button_offset_world_axes_m": { "x": [ -0.3, -0.3 ], "y": [ -0.48, -0.38 ], "z": [ 0.8, 0.8 ] }, "button_quaternion_xyzw": [ 0, -0.70711, 0, 0.70711 ] }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMPushButton.py", "source_sha256": "2bb257ad28b812fd763777879176ceee2796b1e1f3b62a2539bff3558a7c0c1c", "interpretation_caveat": "Normalized joint travel plus finger proximity; no force-contact or sustained-press requirement." }, { "id": "LMPushShelfForward", "instruction": "Push the tall wheeled shelf to the marked area", "public_demonstrations": 100, "matched_checkpoint": "nvidia/G1-Factory-GR00T-1.7", "episode_seconds": 20, "control_hz": 50, "native_success": { "summary": "Shelf/target root separation <[0.3,0.3,0.01]m, static, local-Z/world-up dot \u22650.9, and a finger within1.1m.", "root_separation_lt_xyz_m": [ 0.3, 0.3, 0.01 ], "static": true, "upright_dot_ge": 0.9, "finger_near_m": 1.1 }, "native_reset": { "robot_fixed": true, "cart_center_xyz_m": [ 1.5, 0, -0.8 ], "cart_xy_half_range_m": 0.3, "cart_yaw_center_half_range_rad": [ -1.5707963267948966, 0.06 ], "target_center_xyz_m": [ 4.305, 0, -0.8 ], "target_xy_half_range_m": 0.1, "target_yaw_half_range_rad": 0.02 }, "source_url": "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/environments/LMPushShelfForward.py", "source_sha256": "4a69e60d92a3b25572df83290a9ccc60756e80ae401b84c0135b8194b6d97158", "interpretation_caveat": "Root/target proximity and static/upright test; the1.1m finger proximity allowance is broad." } ], "common_native_metric": { "name": "success_rate", "definition": "Mean native success flag captured before each finishing episode reset; initial reset excluded by SuccessRecorder.", "aggregation": "Report every task separately and equal-task macro average; all task failures remain visible. Pilot seed42 excluded from frozen evaluation seeds1000\u20131009.", "sources": [ "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/tasks/g1_locomanip_lift_task.py", "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/tasks/g1_locomanip_pnp_task.py", "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/tasks/press_button_task.py", "https://github.com/NVlabs/G1-Factory/blob/81cf7909b13fe1a961aae9a4e457a5e786e4df8c/g1_factory/terms/terminations.py", "https://github.com/isaac-sim/IsaacLab-Arena/blob/c5826c87aa181442328c2c74c49f9e551b5c3d8c/isaaclab_arena/metrics/success_rate.py" ] }, "checkpoint_coverage": { "scope": "One joint nine-task GR00T-N1.7 checkpoint; exact published task instructions, G1 WBC31-state/35-command group contract. No demonstrated arbitrary-task language generalization.", "model_card": "https://huggingface.co/nvidia/G1-Factory-GR00T-1.7", "dataset_card": "https://huggingface.co/datasets/nvidia/G1-Factory-Benchmark" }, "risks_and_interpretation": [ "Native lift criteria use absolute object-root world Z, not displacement from reset. Root origins can be below the visible object.", "Approximate grasp means any of14 finger links within0.18m plus at least3 finger joints displaced >0.1rad from defaults; no required object-hand contact sensor, and proximity/closure are not explicitly bound to the same hand.", "Static means linear speed \u22640.15m/s and angular speed \u22640.15rad/s. Placement uses per-axis world root separation, not a support-surface polygon.", "A proximity predicate or a single terminal frame is not proof of a secure sustained grasp. The collection CLI success-hold option does not automatically add dwell to inference success.", "The obstacle task instruction requests avoidance, but its success predicate only tests drill lift/grasp. Keep collisions and videos visible rather than reinterpret native success.", "Button reset_pressedness=0.8 maps through the Arena unpress event to pressed fraction0.2; it does not initialize the button beyond the0.5 success threshold.", "Factory uses an unlocked-waist HOMIEv2 WBC, a distinct robot USD, a different head-camera extrinsic/FOV, and mobile base commands. Apple geometry/controller/camera must not be substituted.", "Official pyproject restricts its uv environment to Linux x86_64. ARM GB300 needs a separately verified runtime; the pinned submodules declare IsaacSim6.0.1 while the model card mentions6.2.", "The pretrained expert was task-tuned on900 public demonstrations plus foundation pretraining. New ACT uses810 gradient-training episodes,90 reserved episodes, task-ID conditioning and320\u00d7240 RGB. Outcomes compare current systems, not architecture causality.", "The90 reserved episodes select ACT checkpoints and calibrate dispersion, so they are training-held-out but not untouched validation. Per-task q95 is empirical score calibration, not failure probability.", "Synchronous RPC pauses simulation. Neither measured wall latency nor native50Hz command timing demonstrates asynchronous latency hiding, Jetson performance, or WAN reliability." ], "viewer_extension_plan": { "storage": "New Factory directory only; previous apple outputs remain immutable.", "default_view": "Task selector + complete paired-seed outcome table, then recorded rollout selection. Do not choose only successful routes.", "recorded_player": "Actual MP4 synchronized by decoded frame presentation timestamp at50Hz; per-step ACT cyan/GR00T purple and five-step amber fade.", "charts": [ "Both wrist mm and degree disagreement", "Both hand rad disagreement", "Waist joint rad diagnostic", "Height m and navigation m/s,rad/s disagreement", "ACT dispersion and task-specific q95", "Critical max ratio and acceptance level1", "Actual selected policy timeline; exclude initial selection from switch count", "RPC wall time and simulation time on separate labeled axes" ], "unknowns": "Expert disagreement absent between queries is a gap, never a zero; measured comparisons are predicted-command disagreement, not tracking error.", "per_task_context": "Instruction, native success predicate and proxy limitations alongside each replay. Locomotion overview must include robot, obstacle and target.", "evidence_links": [ "Exact reset-pair checks", "Native termination predicates and episode record", "USD FK audit", "Frozen model/data/calibration/protocol identities", "All failed attempts and infrastructure retries retained separately" ] } }