ORCS release models

v0.1.1 candidate: adds one joint EgoHTR 03/05 policy. The four v0.1.0 checkpoints are unchanged. Use ORCS branch lok-i/ego-htr while this PR is under review; its manifest pins the exact candidate commit.

Public release checkpoints for ORCS — Oracle Robot Control Synthesis. These are the behavior-validated checkpoints selected from the corresponding private training runs. The W&B project is not required to download or play them.

Play immediately

After installing ORCS and its pinned dependencies:

# Optional: download and verify every checkpoint up front.
bash scripts/setup/download_released_models.sh

# Or let play download only the selected checkpoint on first use.
play Orcs-Dodge-AdaptSonic --agent release --viewer native
play Orcs-PerLoco-Grail-AdaptSonic --agent release --viewer native
play Orcs-PerLoco-OmRe-AdaptSonic --agent release --viewer native
play Orcs-Uolm-AdaptSonic --agent release --viewer native
play Orcs-PerLoco-EgoHtr-AdaptSonic --agent release --num-envs 2 --viewer viser

Downloads are cached by pinned revision under ~/.cache/orcs/releases by default and are verified against the SHA-256 values in release.json. Set ORCS_RELEASE_ROOT to move the cache.

Contents

task	checkpoint	training iteration
Orcs-Dodge-AdaptSonic	Orcs-Dodge-AdaptSonic/checkpoint.pt	7,500
Orcs-PerLoco-Grail-AdaptSonic	Orcs-PerLoco-Grail-AdaptSonic/checkpoint.pt	14,999
Orcs-PerLoco-OmRe-AdaptSonic	Orcs-PerLoco-OmRe-AdaptSonic/checkpoint.pt	10,500
Orcs-PerLoco-EgoHtr-AdaptSonic	Orcs-PerLoco-EgoHtr-AdaptSonic/checkpoint.pt	9,999
Orcs-Uolm-AdaptSonic	Orcs-Uolm-AdaptSonic/checkpoint.pt	19,999

Each .pt file is the complete RSL-RL training checkpoint: actor, critic, optimizer state, iteration, and run metadata. ORCS loads the actor for inference. ONNX exports are intentionally not included in this release.

The PerLoco and UOLM tasks require their matching motion/terrain/object datasets and assets; use the setup workflows in the ORCS repository. The dodge task has no external motion dataset.

Versioning and provenance

The published v0.1.0 tag remains unchanged. This candidate targets v0.1.1; that tag does not exist yet. During review, ORCS pins this PR's immutable commit, not a floating branch or a future tag. release.json records paths, sizes and checksums. EgoHTR has per-model dependency pins; the top-level compatibility record describes the original four policies. Per-task provenance.json files retain training lineage without requiring W&B access.

EgoHTR 03 + 05

• One rank-16 LoRA policy: rocky ascent with hand support and window crossing.
• 10,000 PPO updates; initial-phase annealing ends at update 7,000.
• Ordinary bash scripts/setup/sync_data.sh --no-smpl installs the processed references and colliders. inhouse --no-smpl is a smaller alternative.
• No raw EgoHTR download, UMR or SMPL-X models are needed for playback.
• Paired files: motion PR #14 and assets PR #58.
• These two reviewed scenarios are the demonstrated scope; performance on unreviewed terrain or physical hardware has not been established.

Training references and terrain derive from Alex Brandes and co-authors, ETH Zurich, EgoHTR: Egocentric 4D Demonstrations of Human Terrain Traversal (2026), paper, dataset. The processed data retains CC-BY-4.0. See THIRD_PARTY_NOTICES.md and the task's provenance.json for the source, modifications and original access-conditions summary. This data attribution does not replace the checkpoint's NVIDIA model license.

License

The checkpoints incorporate SONIC model weights and are distributed under the NVIDIA Open Model License. Licensed by NVIDIA Corporation under the NVIDIA Open Model License.

ORCS-authored code and documentation are BSD-3-Clause. Adapted code and all external models, datasets, assets, and body-model files retain their respective terms; see NOTICE, THIRD_PARTY_NOTICES.md, and LICENSES/.

These are special-purpose robotics control models. Validate them in simulation before any use on hardware.