daibfc/h2r-benchmark-feasible43

HOI2ROI feasible43: humanoid walk–pick–carry–place benchmark (Unitree G1)

43 out-of-distribution box loco-manipulation tasks. Each case is a full task in a real indoor room layout (ARKitScenes scans): the person/robot starts at A, walks to a box at B, picks it up, carries it along a planned path and places it at target C. Motions were generated by HOI-FHLI (human–object interaction synthesis) and retargeted to the Unitree G1 (29 DoF) with an interaction-mesh retargeter (holosoma / OmniRetarget). Median length ~15 s (walk-up ~5 s, carry ~2.2 m). The cases are a filtered subset of 140 generated scenes (see "Selection").

Layout

cases/<case_id>/human_motion.npz   generated human motion + box trajectory (30 fps)
cases/<case_id>/g1_retarget.npz    G1 retarget of the same motion (30 fps)
cases/<case_id>/keypoints.json     task keypoints in the room frame: human start A, box B, target C, A* path, floor/ceiling z
cases/<case_id>/box.json           box: name, scale, size (m), initial pose
cases.csv                          case_id, simulator key, scene id, length, hard-case labels
assets/largebox.obj                box mesh (OMOMO largebox)
assets/box_small_0.7601.urdf       the box every motion was generated for (scale 0.7601, ~25.3 x 25.7 x 27.4 cm)
assets/box_large.urdf              the full-size largebox (~33.5 x 33.8 x 36.0 cm)
feasible43.yaml                    case list (simulator keys, per-case hard labels)
sim/                               simulator-ready references in our harness's format (see below)

Formats

human_motion.npz (HOI-FHLI output): root_trans [T,3], poses [T,52,3] (axis-angle; 22 body + 30 hand joints, SMPL-H/X ordering), betas [16], gender, joints [T,24,3], obj_pos [T,3], obj_rot_mat [T,3,3], object_name, fps (30), nav_frames (end of the walk-up), human_scale. The human and the box were generated at scale human_scale = 0.7601 (a ~1.3 m person with the ×0.7601 box, matching the G1's size); positions are stored at that scale. Divide root_trans by human_scale to replay a full-size human.

g1_retarget.npz: qpos [T,43] = root position (3) + root quaternion wxyz (4) + 29 G1 joint positions (MuJoCo order of the Unitree G1 29-DoF model) + box position (3) + box quaternion wxyz (4); human_joints [T,52,3] (the targets the retargeter tracked), fps (30), cost (retarget objective). The G1 is retargeted at robot scale (the 0.7601 human scale is divided out); the box trajectory is kept exactly as generated.

Box frame. All box poses (both npz files, box.json) are poses of the mesh origin of largebox.obj. The small box is the mesh scaled uniformly by 0.7601 about that origin (exactly box_small_0.7601.urdf). Note the mesh axes are rotated relative to the box faces (~26° yaw, ~9° tilt): use the mesh, not its axis-aligned bounding box. Box mass 0.5 kg in both URDFs.

keypoints.json: 2-D room-frame points human, box, target, path (A* waypoints), mid_index, floor_z, ceiling_z.

Simulator-ready version (sim/)

Our Isaac Lab harness's format: teacher_robot.pkl, teacher_objects.pkl, actor_robot.pkl, actor_objects.pkl (joblib dicts keyed by the simulator key sub17_largebox_20xx_original; teacher = actor = the G1 retarget), manifest.json, key_map.json (simulator key → case_id) and plans/<key>.npz (per-frame lift / release flags and planned box orientation in the pelvis frame, 50 Hz, used by a box-pose estimator). In sim/ the box reference was raised so that the full-size box rests on the floor (all published baselines below use the full-size box). To simulate the intended small box from sim/, scale the box by 0.7601 about the centre of its bottom face.

Task and scoring

Episodes start at the first frame and run the whole reference (no early termination). Per episode:

• pickup: the box is lifted above min(30 cm, 50 % of the reference lift height) for ≥ 0.3 s while the robot is upright and a hand touches the box;
• success: pickup AND the final box position is within 30 cm of the goal (the reference's final box position = C).

Selection

From 140 generated scenes, filter53 kept cases a G1 can follow (labels by the team); feasible43 additionally removes grips the G1 cannot copy: palms closer than half the box's narrow side along the pelvis left–right axis at lift (narrow grip; front-to-back squeeze). 5 cases are labelled hard (hard column): a palm on top of a box turned

25° (top_hand_turned_box), a sudden dive onto the box while walking (sudden_pickup), a diagonal near/far-face grip (near_far_diagonal_grip). Report all 43 and the 38 non-hard cases.

Baselines (full-size box, 43 cases × 8 episodes, no steering)

policy	pickup	success
R15 (residual over frozen SONIC, rubber hands, estimated box pose)	29.4 %	20.1 %
P23 (residual over frozen SONIC, Dex3 hands, true box pose)	20.3 %	11.3 %

Policies: daibfc/h2r-weights. Results on the intended ×0.7601 box will be added.

Sources and licenses

Human motions are generated by HOI-FHLI (trained on OMOMO); the room layouts come from ARKitScenes (room meshes are not included); largebox.obj is the OMOMO largebox. Follow the licenses of OMOMO, ARKitScenes and the SMPL family models when using this data. No real human capture is included.