Assets and checkpoints

Bundled

The repository includes one Habitat split header plus 41 scene shards. The header defines the six category mappings; the 41 shards contain all 1,218 Find/Nav/Interact episodes. It also includes the beta-zero hand-merged runtime body, an optional finger-separated beta-zero body, natural seed state, stabilized environment metadata, and provenance records. The hand-merged body remains the paper default. The optional body gives all 30 finger links separate visual and collision meshes and must be selected explicitly with --agent-asset finger-separated; its added contacts can change physics and collision metrics. It also includes the 66 pelvis-relative neutral-joint constants used by Motion Jerk in resources/metrics/smpl_neutral_body22.json; the full SMPL archive is neither needed nor redistributed. Versioned planner and verifier prompts are ordinary, inspectable Python modules under agent/. Precomputed model results are not bundled.

resources/benchmark/val100.json is an episode index for the historical five-scene, 100-episode development subset. It contains no duplicate scene or episode payloads; every row points into the same canonical full-validation shards. It was recovered from the archived val_s5_e20_seed20260530.json manifest (source SHA-256 6182cc3a224353e34a54bd0946fd7a9639eb9bc4e0d6278813bde84cfc5a43b7); all 100 scene/episode/category rows and the seed/category counts match that source exactly. Select it with humanclaw-bench run --episodes val100.

Every episode stores start_position, start_rotation, init_offset, and init_yaw directly. The provenance file spawn_repair_history_20260806_v2.csv lists which validation pass supplied the final humanoid coordinates for each of the 308 corrected episodes. It is not a runtime input. See SPAWN_REPAIRS.md for the merge rules.

Run humanclaw-bench assets to verify their deterministic file or tree hashes. The verifier checks file count and total size as well as SHA-256 for directory assets.

External HSSD data

Official HSSD stage and object meshes remain external because they have their own access and redistribution terms. The full-val benchmark references 41 scenes. Prepare an authorized HSSD val installation with:

humanclaw-bench prepare-hssd --hssd-root /path/to/hssd-hab

resources/hssd/humanclaw-hssd-val41/ contains the 41 scene descriptions, 14,537 directly inspectable per-instance physics configs, and the immutable manifest for 1,693 instance-specific baked meshes absent from the official download. The meshes are distributed as a gated 79.8 MiB Hugging Face asset; their extracted logical size is 176 MiB. asset_requirements.json pins every official and supplemental asset by size and SHA-256. The generated data/humanclaw-hssd-val41/ reuses the official tree and verified HF cache through symlinks. The supplement is required for baked scale fixes and exact render-mesh collision geometry.

The default prepare-hssd command downloads the pinned archive automatically. For an offline host, pass the downloaded archive with --supplement.

Pass --output to prepare elsewhere, then pass its hssd-hab.scene_dataset_config.json to run with --scene-dataset-config.

External motion data

AMASS, BABEL, and the neutral SMPL model are not redistributed. The repository contains only eight transparent CSV training lists under resources/motion/training/manifests/ (1.67 MB total). Each row identifies a relative AMASS chunk pickle and its source index; it does not contain motion arrays. A second set under resources/motion/training/segments/ reports the human-reviewed BABEL interval and the first/last final-used chunk time in both seconds and milliseconds. Its per-segment chunk counts sum exactly to the machine training lists. Build the 20-frame corpus locally and use these lists as described in the motion-training guide.

External motion weights

resources/weights/paper_fullval_v1.json is authoritative. It pins one canonical MotionDiT state, two deterministic numerical base variants, and eight control-only skill states selected from exact step 1,500,000. There is no latest resolution.

The nine inference files are distributed as HumanCLAW_pretrained_weights_paper_fullval_v1_20260816.tar.gz in the HumanCLAW/HumanCLAW Hugging Face model repository. The archive SHA-256 is 3b3c0c1b232af4c462301655de909a4bc54fd4756bcd42c22e7965fccb650667.

Skill	Network	Conditioning
walk_forward	WalkForwardCtrlDiTFourierXZYaw	Fourier x/z/yaw, 6 frequencies
side_walk	SideWalkCtrlDiTFourier	Fourier, 6 frequencies
step_back	WalkForwardCtrlDiTFourier	Fourier, 6 frequencies
turn	TurnCtrlDiT	MLP
step_climb_up	StepClimbUpCtrlDiT	MLP
step_climb_down	StepClimbDownCtrlDiT	MLP
stop	NonCondCtrlDiT	none
sit	SitCtrlDiT	MLP

The base architecture is 219 input/output dimensions, hidden size 512, 10 layers, 8 heads, MLP ratio 2, five history frames, 15 future frames, and 30 flow-evaluation steps. walk_forward, turn, stop, and sit use the exact BF16-roundtrip variant embedded in their archived checkpoints; the other four use the canonical FP32 tensors. The former is a deterministic rounded copy of the latter, not a separately trained MotionDiT. The weight-directory README documents the audit and why retaining both numerical variants matches the reported evaluation.

The original nine files occupied 3,227,420,338 bytes because every ControlNet repeated its frozen base and the base Lightning checkpoint retained training state. The inference-only set is 1,135,639,268 bytes. The loader recreates all eight archived model_state mappings tensor-for-tensor and shares each base variant among its four skills in memory.

Checkpoint validation occurs before any checkpoint is unpickled. This protects reproducibility and reduces accidental loading of an unexpected artifact; it does not make untrusted pickle files safe. Obtain weights only from the maintainer-approved distribution.

Redistribution gate

Do not upload scene data, body assets, or weights merely because they appear in a local working copy. Confirm third-party terms and add notices first. The release checklist treats this as a publication blocker.