Bagel-Robotics-1
Bagel-Robotics-1 is an open-weight world-action model for two-arm robot manipulation from Bagel Labs, a physical AI research lab based in Toronto. It predicts arm and gripper actions from a task instruction, three camera views, and robot state. At its launch, it ranked No. 3 worldwide among standalone models by average score on RoboDojo-Sim.
How did it score on RoboDojo-Sim?
Bagel Labs identified Bagel-Robotics-1 as the highest-ranked model from a North American company in its launch announcement.
- Average score
- 38.65 / 100
- Average task success rate
- 31.91%
The leaderboard ranks by partial-credit score. The average gives equal weight to generalization, precision, long-horizon tasks, memory, and open tasks. These results are from simulation with the evaluated checkpoint. The leaderboard shows current standings.
Anthropic ran RoboDojo-Sim internally for its Claude Fable 5.2 evaluation. Bagel’s ranking uses RoboDojo’s standalone model track, which reports agent entries separately.
How was Bagel-Robotics-1 trained?
Bagel Labs used Data-Efficient Adaptive Pruning (DEAP) to select demonstration windows and prune and train the video expert while keeping the action expert fixed. The video expert supplies visual features to the action expert. The published method explains the training and export.
The public release includes the evaluated cp1000 checkpoint and inference files. The model card identifies the bundle and documents the simulator and additional assets needed to reproduce the evaluation.
What does the license allow?
The Bagel Labs Research License allows eligible researchers and institutions to use Bagel’s contributions for noncommercial research. For-profit and commercial use require a separate agreement. Upstream components retain their own terms.
For a robotics use case, register deployment interest or contact Bagel Labs about a partnership.