Bagel Labs

Bagel-Robotics-1

Bagel-Robotics-1 is Bagel Labs’ open-weight world-action model for bimanual robot manipulation. Given a task instruction, three camera views and the robot’s current state, it predicts how to move both arms and grippers.

As of November 20, 2026, Bagel Labs’ Bagel-Robotics-1 ranked third among standalone models on RoboDojo-Sim.

GitHub reference implementation RoboDojo-Sim leaderboard Bagel Labs robot demos

RoboDojo-Sim Rank Ranking metric
Generalization Standard 🥇 Tied 1st 46.44% success
Memory 🥈 2nd 57.21/100 partial score
Overall 🥉 3rd Average partial score across five dimensions

Watch and run the two-arm tower example · Browse all 54 RoboDojo-Sim task configurations

Download the checkpoint · Run inference · Read the method · Licensing

Two-arm robot manipulation with Bagel-Robotics-1 by Bagel Labs

In RoboDojo-Sim, Bagel-Robotics-1 uses both arms of an ARX X5 to build a tower. The grippers lift a wooden board together and place it across two supports. RoboDojo recorded this successful build_tower episode with checkpoint cp1000, seed 0 and layout 0.

Watch the official recording (29 seconds) · Cite the Bagel Labs model release

Run one episode. Follow the adapter and asset setup on Linux with a CUDA GPU, using RoboDojo f57e163 and XPolicyLab 6baccd5. From the RoboDojo root, run the command below. Replace --eval-env RoboDojo if your simulator environment has a different name.

BAGEL_POLICY_DIR="$PWD/XPolicyLab/policy/Bagel_Robotics_1"
EVAL_ENV_TYPE=sim bash scripts/robodojo.sh eval \
  --policy-dir "$BAGEL_POLICY_DIR" --policy-env "$BAGEL_POLICY_DIR/.venv" \
  --eval-env RoboDojo --task build_tower --ckpt cp1000 \
  --env-cfg arx_x5 --action-type ee --seed 0 --eval-num 1

RoboDojo writes the outcome to _result.json in the run directory under eval_result/RoboDojo/build_tower/. Append --dry-run to preview the command before loading the model.

Bagel Labs robot demos

✨ Overview

The policy uses images from a head camera and two wrist cameras. It predicts 32 actions at a time and executes 16 before replanning from fresh observations.

What did Bagel Labs train?

Bagel Labs trained the video expert on robot demonstrations while keeping the action expert fixed. The video expert supplies the visual features used to predict both arms’ movements. The training data spans 35 tasks in the RoboDojo dataset, with 30,000 demonstration windows selected for fitting and 1,119 for monitoring.

Training method and architecture

Bagel Labs’ Data-Efficient Adaptive Pruning (DEAP) method selects demonstration windows, learns which components of the video expert to retain, and trains the resulting expert with action supervision. This reduced the video expert from 16,394,878,784 to 12,468,409,664 parameters, a 23.95% reduction.

DEAP training stages and the deployed policy’s action loop

The export uses smaller FFN and cross-attention layers and retains the self-attention heads. Inference runs directly from the exported architecture and weights.

📊 Official RoboDojo-Sim results

Official RoboDojo-Sim leaderboard results, as of November 20, 2026, excluding the Agent track:

Dimension Partial score /100 Success rate (%) Ranking highlight
Overall average 38.65 31.91 🥉 3rd by partial score
Average without Open 47.33 39.05 —
Generalization 35.52 28.56 —
Generalization Standard (Gen-Std) 52.45 46.44 🥈 2nd by partial score · 🥇 Tied 1st by success rate
Generalization Random (Gen-Rand) 18.59 10.67 —
Precision 42.65 35.42 —
Long-Horizon 53.94 35.33 —
Memory 57.21 56.89 🥈 2nd by partial score
Open 3.93 3.33 —

The overall average weights Generalization, Precision, Long-Horizon, Memory and Open equally. The average without Open uses the first four dimensions. Rankings are a dated leaderboard snapshot; refer to the official source for current standings.

See the results summary for machine-readable scores and rankings, and the evaluation guide for settings and comparison requirements.

Fixed checkpoint identity

  • Download checkpoint 1000 at release tag cp1000.
  • Identity: 790bffc9464fa7a026aa43b217aec2033d186895cad643166daa0bcc01914d5c.
  • artifact-manifest.json pins all six bundle files and the separate public encoder/tokenizer assets.
  • SHA256SUMS records the unchanged SHA-256 hashes of all six bundle files.

COMPLETE.json records the original export. All six checkpoint files match the ones used for the reported evaluation.

Research and commercial use

Independent researchers and teams at universities, nonprofits, and public research institutions can use Bagel Labs’ contributions to the weights free of charge for noncommercial research. You can run the model, fine-tune it on your own robot demonstrations, and publish your results under the Bagel Labs Research License 1.0.

Commercial use and all use by or on behalf of a for-profit company, including internal research and benchmarking, require a separate paid agreement. Contact Bagel Labs about licensing.

Download the exact checkpoint

The checkpoint and inference files are public and ungated. You can download them without a Hugging Face account or access token.

hf download bageldotcom/Bagel-Robotics-1 LICENSE.md THIRD_PARTY_NOTICES.md \
  --local-dir /absolute/path/to/checkpoints/cp1000

hf download bageldotcom/Bagel-Robotics-1 \
  --revision cp1000 \
  --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.

Inference adapter and reevaluation

The standalone XPolicyLab adapter is provided under inference/. The current public adapter follows the instructions in inference/README.md; no private research repository or access token is required.

hf download bageldotcom/Bagel-Robotics-1 \
  --include 'inference/**' \
  --local-dir /absolute/path/to/Bagel-Robotics-1-release

# Then install the adapter into an existing XPolicyLab checkout.
python /absolute/path/to/Bagel-Robotics-1-release/inference/install_adapter.py /absolute/path/to/workspace/XPolicyLab

Follow inference/policy/Bagel_Robotics_1/README.md to install the Python 3.10 environment, download and verify assets, and run both debug checks. Follow REEVALUATION.md for task, suite and three-seed commands. Install RoboDojo and simulator assets separately using its official guide.

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.

Validation and release scope

The adapter qualification includes 15 CPU tests, both real-checkpoint debug transports, four valid native simulator episodes and eight exact action/history/reset/backend checks on tested recordings.

The training and export code is available in the Bagel-Robotics-1 GitHub repository. This Hugging Face repository contains the inference release and weights; it does not claim that optimizer state or exact training continuation can be reconstructed from weights alone.

Citation

@software{bagellabs2026bagelrobotics1,
  author  = {{Bagel Labs}},
  title   = {{Bagel-Robotics-1}},
  year    = {2026},
  version = {cp1000},
  url     = {https://huggingface.co/bageldotcom/Bagel-Robotics-1}
}

Machine-readable citation

Attribution

See the source and artifact notices for upstream credits, source provenance and license terms. Thanks to the RoboDojo and XPolicyLab teams for the benchmark and evaluation framework.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Dataset used to train bageldotcom/Bagel-Robotics-1