File size: 7,344 Bytes
45f4f61
 
 
 
 
 
df9c339
 
c436577
45f4f61
9c600ba
45f4f61
4127bc7
b6798ea
4127bc7
 
 
 
 
45f4f61
4127bc7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45f4f61
9c600ba
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bdb727b
45f4f61
bdb727b
 
7900dfe
 
45f4f61
7900dfe
45f4f61
bdb727b
45f4f61
4127bc7
 
bdb727b
12bfd47
bdb727b
 
 
 
45f4f61
7900dfe
45f4f61
bdb727b
9b01ecf
7900dfe
9b01ecf
 
12bfd47
bdb727b
12bfd47
9b01ecf
7900dfe
12bfd47
bdb727b
9b01ecf
7900dfe
45f4f61
7900dfe
45f4f61
bdb727b
45f4f61
9c600ba
45f4f61
9c600ba
45f4f61
7900dfe
45f4f61
4127bc7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
---
language: en
tags:
- robotics
- structured-pruning
- robodojo
- '#physicalai'
- '#wam'
---

# Bagel-Robotics-1

An open-weight world-action model for robot manipulation from Bagel Labs.

<p align="center">
  <a href="https://github.com/bageldotcom/Bagel-Robotics-1"><img alt="GitHub reference implementation" src="https://img.shields.io/badge/GitHub-Reference_Implementation-202936"></a>
  <a href="https://robodojo-benchmark.com/leaderboard?bench=sim"><img alt="RoboDojo-Sim leaderboard" src="https://img.shields.io/badge/RoboDojo--Sim-Leaderboard-365F98"></a>
  <a href="https://www.bagel.com/#real-world-demo"><img alt="Bagel Labs robot demos" src="https://img.shields.io/badge/Bagel_Labs-Robot_Demos-3B6055"></a>
</p>

| πŸ₯‰ RoboDojo-Sim | βœ‚οΈ Compact video expert | πŸ“¦ Public release |
| :---: | :---: | :---: |
| **3rd overall** among standalone models by partial score | **12.468B parameters**, down **23.95%** | Checkpoint weights and inference code |

*RoboDojo-Sim ranking as of November 20, 2026.*

[Download the checkpoint](#download-the-exact-checkpoint) Β· [Run inference](#inference-adapter-and-reevaluation) Β· [Read the method](https://github.com/bageldotcom/Bagel-Robotics-1/blob/main/docs/METHOD.md)

## 🎬 Robot demos

| Coffee preparation | Tube placement |
| :---: | :---: |
| [![Robot arm holding a cup by a coffee machine](assets/robot-demo.webp)](https://www.bagel.com/#real-world-demo) | [![Robot arm placing a tube in a rack, shown from three camera angles](assets/robot-demo-tubes.webp)](https://www.bagel.com/#tube-placement-demo) |

[Watch both demos on the Bagel Labs website](https://www.bagel.com/#real-world-demo).

## ✨ Overview

Bagel-Robotics-1 uses **Data-Efficient Adaptive Pruning (DEAP)** to select informative windows from robot demonstrations, prune the video expert, and recover action quality by training the smaller model.

![DEAP pipeline: select demonstrations, learn masks, prune the video expert, and run recovery training.](assets/approach.svg)

The video expert has **12,468,409,664 parameters**, 23.95% fewer than its pretrained starting point. Learned masks were converted into smaller FFN and cross-attention layers. Self-attention heads and the action expert are retained. Inference uses the exported model directly and requires neither teacher weights nor a learned-mask module.

## πŸ“Š Official RoboDojo-Sim results

Official [RoboDojo-Sim leaderboard](https://robodojo-benchmark.com/leaderboard?bench=sim) results, **as of November 20, 2026**, excluding the Agent track:

| Dimension | Partial score /100 | Success rate (%) | Ranking highlight |
| --- | ---: | ---: | --- |
| **Overall average** | **38.65** | **31.91** | πŸ₯‰ **3rd** by partial score |
| Average without Open | 47.33 | 39.05 | β€” |
| Generalization | 35.52 | 28.56 | β€” |
| **Generalization Standard (Gen-Std)** | **52.45** | **46.44** | πŸ₯ˆ **2nd** by partial score Β· πŸ₯‡ **Tied 1st** by success rate |
| Generalization Random (Gen-Rand) | 18.59 | 10.67 | β€” |
| Precision | 42.65 | 35.42 | β€” |
| Long-Horizon | 53.94 | 35.33 | β€” |
| **Memory** | **57.21** | **56.89** | πŸ₯ˆ **2nd** by partial score |
| Open | 3.93 | 3.33 | β€” |

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](https://github.com/bageldotcom/Bagel-Robotics-1/blob/main/results/summary.json) for machine-readable scores and rankings, and the [evaluation guide](https://github.com/bageldotcom/Bagel-Robotics-1/blob/main/docs/EVALUATION.md#reported-results) for settings and comparison requirements.

## 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. `COMPLETE.json` is the original export receipt, retained unchanged for integrity and not used as a current evaluation-status claim.

## 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.

```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.

## Inference adapter and reevaluation

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

```bash
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`](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`](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.

## 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, pruning and export implementation is publicly available in the [Bagel-Robotics-1 GitHub repository](https://github.com/bageldotcom/Bagel-Robotics-1). The technical report and citation are forthcoming. 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.

## Attribution

See the [source and artifact notices](THIRD_PARTY_NOTICES.md) for upstream credits, source provenance and license terms. Thanks to the RoboDojo and XPolicyLab teams for the benchmark and evaluation framework.