Rename policy directory to Bagel_Robotics_1

004286e verified 2 days ago

6.22 kB

---
language: en
tags:
- robotics
- vpp2
- structured-pruning
- robodojo
---
# Bagel-Robotics-1 — checkpoint 1000

Bagel-Robotics-1 combines mask-based structured pruning, recovery training and task-critical data selection to produce a compact WAM with nonuniform FFN widths and cross-attention head counts. Its inference pipeline predicts 32 actions, executes 16 before replanning, and applies 90/10 temporal ensembling. VPP2 is the validation base; reused source components are credited in the notices below.

Inference release of the **exact checkpoint 1000** used in the Bagel-Robotics-1 evaluation campaign. **The repository is currently private.** Evaluator access is being arranged using a dedicated fine-grained token with read access only to this repository, delivered through an agreed private channel. The dedicated repository-only read credential is verified; assigned evaluator handoff remains pending. Public release is prepared for the later verified-leaderboard milestone and requires a bageldotcom administrator.

The video expert has **12,468,409,664 parameters**, 23.95% fewer than its recorded parent. Learned masks were converted into smaller FFN and cross-attention layers. Self-attention heads and the action expert are retained. No teacher or learned-mask module is required for inference. Parameter reduction is not a measured inference-speedup claim.

## Fixed checkpoint identity

- Original checkpoint revision: `2651631c5d08819088475e8778676f5cf6d9afcf`.
- Identity: `790bffc9464fa7a026aa43b217aec2033d186895cad643166daa0bcc01914d5c`.
- [artifact-manifest.json](artifact-manifest.json) pins all six bundle files and the separate public encoder/tokenizer assets.
- [SHA256SUMS](SHA256SUMS) records the unchanged SHA-256 hashes of all six bundle files.

The release adds documentation and inference packaging without changing any checkpoint file. The original revision is preserved, so historical provenance remains accessible in Git history. Current release files focus on inference. `COMPLETE.json` is the original export receipt, including export-time paths and `simulation_completed: false`; it is retained unchanged for integrity, not used as a current evaluation-status claim. Those historical paths are not required for deployment.

## Download the exact checkpoint

While the repository is private, supply your authorized evaluator credential securely through `HF_TOKEN` or your Hugging Face login. Do not put credentials in commands, public documents, or application text. The commands below require that access.

```bash
hf download bageldotcom/Bagel-Robotics-1 \
  --revision 2651631c5d08819088475e8778676f5cf6d9afcf \
  --include video.pt action.pt compact_architecture.json dataset_stats.json COMPLETE.json validation.json \
  --local-dir /absolute/path/to/checkpoints/cp1000
```

There are approximately 29.0 GB of checkpoint tensors. Keep all six files together. The inference downloader checks their sizes and SHA-256 hashes and downloads approximately 16.7 GB of separately pinned VAE/conditioning/tokenizer assets. Allow extra disk space for caches and installation.

## Inference adapter and reevaluation

The standalone XPolicyLab adapter is provided under [`inference/`](inference/), at release tag `bagel-robotics-1-cp1000-inference-v2`. No private research repository is required.

```bash
hf download bageldotcom/Bagel-Robotics-1 \
  --revision bagel-robotics-1-cp1000-inference-v2 \
  --include 'inference/**' \
  --local-dir /absolute/path/to/Bagel-Robotics-1-release
git clone https://github.com/XPolicyLab/XPolicyLab.git /absolute/path/to/workspace/XPolicyLab
git -C /absolute/path/to/workspace/XPolicyLab checkout 6baccd57ff958570d0db1588370289ee36c8bb70
python /absolute/path/to/Bagel-Robotics-1-release/inference/install_adapter.py /absolute/path/to/workspace/XPolicyLab
```

Follow the included [policy README](inference/policy/Bagel_Robotics_1/README.md) to install the Python 3.10 environment, download/verify assets and run both debug checks. Follow [REEVALUATION.md](inference/policy/Bagel_Robotics_1/REEVALUATION.md) for task, suite and three-seed commands. Install RoboDojo and simulator assets separately using its [official guide](https://robodojo-benchmark.com/doc/usage/install-and-download/).

The adapter uses `policy_name=Bagel_Robotics_1`, `ckpt_name=cp1000`, ARX X5 end-effector actions, and one environment per policy process. Fixed inference settings: BF16; predict 32 actions and execute 16; ten Euler steps; shift 1; 90% new / 10% previous temporal blending. Qualification used a 96 GB RTX PRO 6000 Blackwell, Python 3.10.21 and PyTorch 2.11.0+cu130; a lower minimum GPU capacity has not been qualified.

Use the Bagel-Robotics-1 compact checkpoint loader supplied in the inference package. The implementation and current reviewer instructions are available in [XPolicyLab PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234).

## Validation and release scope

The adapter's existing qualification includes 15 CPU tests, both real-checkpoint debug transports, four valid native simulator episodes, and eight exact action/history/reset/backend checks on the tested recordings. Full multi-seed official leaderboard validation remains separate. Historical research scores and scene exceptions are documented in the adapter's EVALUATION_RESULTS.md; they are not official leaderboard results.
The technical report and training/pruning/export code are planned for **November 20, 2026**. This release contains inference code and weights; optimizer, scheduler, distributed RNG/sampler state, training datasets and simulator assets are not included. It does not support exact training continuation from weights alone.
## Attribution and terms
VPP2 is the base model used to validate the Bagel-Robotics-1 pruning, recovery-training and data-selection approach. Reused source components come from [VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM and Wan2.1/DiffSynth. See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the adapter's retained MIT/Apache-2.0 notices. Model and auxiliary-asset upstream terms continue to apply; an inference-code license is not a new blanket license for all weights or datasets.

	---
	language: en
	tags:
	- robotics
	- vpp2
	- structured-pruning
	- robodojo
	---

	# Bagel-Robotics-1 — checkpoint 1000

	Bagel-Robotics-1 combines mask-based structured pruning, recovery training and task-critical data selection to produce a compact WAM with nonuniform FFN widths and cross-attention head counts. Its inference pipeline predicts 32 actions, executes 16 before replanning, and applies 90/10 temporal ensembling. VPP2 is the validation base; reused source components are credited in the notices below.

	Inference release of the exact checkpoint 1000 used in the Bagel-Robotics-1 evaluation campaign. The repository is currently private. Evaluator access is being arranged using a dedicated fine-grained token with read access only to this repository, delivered through an agreed private channel. The dedicated repository-only read credential is verified; assigned evaluator handoff remains pending. Public release is prepared for the later verified-leaderboard milestone and requires a bageldotcom administrator.

	The video expert has 12,468,409,664 parameters, 23.95% fewer than its recorded parent. Learned masks were converted into smaller FFN and cross-attention layers. Self-attention heads and the action expert are retained. No teacher or learned-mask module is required for inference. Parameter reduction is not a measured inference-speedup claim.

	## Fixed checkpoint identity

	- Original checkpoint revision: `2651631c5d08819088475e8778676f5cf6d9afcf`.
	- Identity: `790bffc9464fa7a026aa43b217aec2033d186895cad643166daa0bcc01914d5c`.
	- [artifact-manifest.json](artifact-manifest.json) pins all six bundle files and the separate public encoder/tokenizer assets.
	- [SHA256SUMS](SHA256SUMS) records the unchanged SHA-256 hashes of all six bundle files.

	The release adds documentation and inference packaging without changing any checkpoint file. The original revision is preserved, so historical provenance remains accessible in Git history. Current release files focus on inference. `COMPLETE.json` is the original export receipt, including export-time paths and `simulation_completed: false`; it is retained unchanged for integrity, not used as a current evaluation-status claim. Those historical paths are not required for deployment.

	## Download the exact checkpoint

	While the repository is private, supply your authorized evaluator credential securely through `HF_TOKEN` or your Hugging Face login. Do not put credentials in commands, public documents, or application text. The commands below require that access.

	```bash
	hf download bageldotcom/Bagel-Robotics-1 \
	--revision 2651631c5d08819088475e8778676f5cf6d9afcf \
	--include video.pt action.pt compact_architecture.json dataset_stats.json COMPLETE.json validation.json \
	--local-dir /absolute/path/to/checkpoints/cp1000
	```

	There are approximately 29.0 GB of checkpoint tensors. Keep all six files together. The inference downloader checks their sizes and SHA-256 hashes and downloads approximately 16.7 GB of separately pinned VAE/conditioning/tokenizer assets. Allow extra disk space for caches and installation.

	## Inference adapter and reevaluation

	The standalone XPolicyLab adapter is provided under [`inference/`](inference/), at release tag `bagel-robotics-1-cp1000-inference-v2`. No private research repository is required.

	```bash
	hf download bageldotcom/Bagel-Robotics-1 \
	--revision bagel-robotics-1-cp1000-inference-v2 \
	--include 'inference/**' \
	--local-dir /absolute/path/to/Bagel-Robotics-1-release

	git clone https://github.com/XPolicyLab/XPolicyLab.git /absolute/path/to/workspace/XPolicyLab
	git -C /absolute/path/to/workspace/XPolicyLab checkout 6baccd57ff958570d0db1588370289ee36c8bb70
	python /absolute/path/to/Bagel-Robotics-1-release/inference/install_adapter.py /absolute/path/to/workspace/XPolicyLab
	```

	Follow the included [policy README](inference/policy/Bagel_Robotics_1/README.md) to install the Python 3.10 environment, download/verify assets and run both debug checks. Follow [REEVALUATION.md](inference/policy/Bagel_Robotics_1/REEVALUATION.md) for task, suite and three-seed commands. Install RoboDojo and simulator assets separately using its [official guide](https://robodojo-benchmark.com/doc/usage/install-and-download/).

	The adapter uses `policy_name=Bagel_Robotics_1`, `ckpt_name=cp1000`, ARX X5 end-effector actions, and one environment per policy process. Fixed inference settings: BF16; predict 32 actions and execute 16; ten Euler steps; shift 1; 90% new / 10% previous temporal blending. Qualification used a 96 GB RTX PRO 6000 Blackwell, Python 3.10.21 and PyTorch 2.11.0+cu130; a lower minimum GPU capacity has not been qualified.

	Use the Bagel-Robotics-1 compact checkpoint loader supplied in the inference package. The implementation and current reviewer instructions are available in [XPolicyLab PR #234](https://github.com/XPolicyLab/XPolicyLab/pull/234).

	## Validation and release scope

	The adapter's existing qualification includes 15 CPU tests, both real-checkpoint debug transports, four valid native simulator episodes, and eight exact action/history/reset/backend checks on the tested recordings. Full multi-seed official leaderboard validation remains separate. Historical research scores and scene exceptions are documented in the adapter's EVALUATION_RESULTS.md; they are not official leaderboard results.

	The technical report and training/pruning/export code are planned for November 20, 2026. This release contains inference code and weights; optimizer, scheduler, distributed RNG/sampler state, training datasets and simulator assets are not included. It does not support exact training continuation from weights alone.

	## Attribution and terms

	VPP2 is the base model used to validate the Bagel-Robotics-1 pruning, recovery-training and data-selection approach. Reused source components come from [VPP2](https://github.com/roboterax/video-prediction-policy-2)/FastWAM and Wan2.1/DiffSynth. See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the adapter's retained MIT/Apache-2.0 notices. Model and auxiliary-asset upstream terms continue to apply; an inference-code license is not a new blanket license for all weights or datasets.