Bagel-Robotics-1 Takes No. 3 Spot on RoboDojo Benchmark

In the high-stakes playground of physical AI, the standard playbook has long been simple: if your robot drops the coffee mug, throw another fifty billion parameters and three megawatt-hours of compute at the problem. Silicon Valley hyperscalers have spent the last three years insisting that general-purpose robotics is purely downstream of gargantuan Vision-Language-Action (VLA) foundation models.

Then came Toronto-based Bagel Labs, serving up a surprisingly crispy counterpoint.

In its November launch, the Canadian lab’s open-weight Bagel-Robotics-1 reached the No. 3 spot among standalone robot-control policies on RoboDojo-Sim, ranked by partial-credit score. The Bagel Labs release compared its results with published scores for large commercial systems, including OpenAI’s GPT-6 Astra. The benchmark also has a separate agent track, so third place comes with a category label.

Making the achievement all the sweeter for the robotics community, Bagel Labs did not lock their creation behind a gated enterprise API with a waiting list. They released it as an open-weight model, complete with training pipelines and inference scripts.

The Crucible: What Makes RoboDojo the Ultimate Test?

To understand why third place is causing such a stir in AI laboratories, one must look at the proving ground itself. Unlike static computer vision datasets where models merely label bounding boxes or guess the next pixel in an idealized video clip, RoboDojo is designed to evaluate generalist bimanual robotic manipulation in unforgiving, closed-loop environments.

As detailed in the comprehensive RoboDojo paper on arXiv, the benchmark tests whether an embodied agent can actually survive contact with reality. It does not just measure raw arm trajectory accuracy; it actively grades a policy on five gruelling performance dimensions:

  • Precision: Fine motor tasks demanding sub-millimeter clearances and delicate force interaction.
  • Memory: The capacity to track occluded objects and remember environmental state across multi-step procedures.
  • Generalization: Adapting to novel object geometries, varying friction coefficients, and unseen lighting conditions.
  • Long Horizon: Complex, multi-stage sequences where a single intermediate failure dooms the entire rollout.
  • Open-Ended Interaction: Handling unscripted instructions and ambiguous spatial goals without hardcoded safety nets.

RoboDojo has quickly emerged as the gold standard for embodied intelligence, even earning a dedicated evaluation spotlight in recent frontier foundational system cards. Simply put: you cannot hallucinate your way to a high score on RoboDojo. If your robotic policy gets confused about spatial depth, the virtual gripper smashes the teacup, and your benchmark score plunges into the single digits.

Dissecting the Scorecard: What Third Place Means

The official launch result was 38.65 points out of 100 and 31.91% full task success. A policy can win partial credit for completing parts of a task, so these two numbers measure different things. The benchmark’s open-ended suite was much harder for Bagel: its full success rate there was 3.33%. The leaderboard is a simulation result and can change as new systems are evaluated.

Under the Hood: A Pruned Video Expert

Bagel Labs started from VPP2 and kept the action expert fixed while retraining the video expert on selected robot demonstrations. Its technical release reports that pruning reduced the video expert from 16,394,878,784 to 12,468,409,664 parameters, a reduction of 23.95% in that component. This figure does not establish an equivalent improvement in inference speed.

The policy takes an instruction, a head-camera image, two wrist-camera views, and robot state. It predicts 32 actions and runs 16 before using new observations to plan again.

Open Weights for the Physical Edge

Beyond benchmark bragging rights, the most consequential aspect of Bagel-Robotics-1 is its accessibility. The robotics field has suffered under an epidemic of “look at our cool video, but no, you cannot inspect the weights” announcements.

Bagel Labs has taken the opposite path. Roboticists, academic researchers, and industrial automation engineers can download the model directly: inspect the model repository on Bagel-Robotics-1 on Hugging Face or explore the training harness and RoboDojo evaluation adapters via Bagel-Robotics-1 on GitHub.

The weights are available under the Bagel Labs Research License for eligible noncommercial research. The repository gives the separate terms for commercial and for-profit use.

The Robot King’s Take: The Edge Strikes Back

There is a delicious irony in watching a model named after breakfast fare climb a robotics leaderboard. The interesting result is how much a policy can gain from choosing the right demonstrations and pruning a specific part of the model. The next test is latency and reliability on physical robots, especially when someone hands the robot a task the benchmark did not rehearse.