Bagel-Robotics-1 Beats GPT-6 Astra on Four RoboDojo Suites, Trails 8x on Novel-Task Generalization
Structured pruning cut 24% of parameters but may have cost the generalization capacity deployment demands

Bagel Labs released Bagel-Robotics-1 on November 18, 2026 — a robot manipulation model that ranks third overall on the RoboDojo benchmark and beats OpenAI's GPT-6 Astra on four of the benchmark's five task suites. The fifth suite, Open, measures how well a model handles tasks withheld from submitters before evaluation, and on that suite the gap runs in the opposite direction: Bagel-Robotics-1 scores 4.86/3.67% against GPT-6 Astra's 34.36/31.00% — roughly an 8.5x deficit on the dimension most predictive of real-world deployment value.
The company's press release led with the wins and did not mention the Open suite result.
What Bagel Labs Released, and What the Press Release Left Out
Bagel-Robotics-1 is a Fast-WAM architecture model, meaning it belongs to the World-Action Model class — systems that jointly learn future visual prediction and robot action generation during training. At inference time, Fast-WAM skips the future-prediction step and runs a single forward pass, which reduces latency while retaining the representational benefits of joint training.
Bagel Labs applied adaptive structured pruning to the video expert component of its base model. Unlike unstructured pruning, which removes individual weights and produces irregular sparse matrices, structured pruning removes entire channels, attention heads, or feed-forward network units. The result is a model that runs on standard hardware without specialized sparse-inference libraries. Bagel's video expert was reduced from approximately 15.0B to 12.47B parameters — a 23.95% reduction. After pruning, the team ran a recovery training pass to restore accuracy.
The final Bagel-Robotics-1 model is open-weight and published on HuggingFace.
RoboDojo, presented at ECCV 2026, is a unified simulation-and-real benchmark covering five suites: Generalization, Precision, Long-Horizon, Memory, and Open. The first four suites are disclosed to submitters before evaluation. The Open suite is not: tasks are withheld until the model is submitted, specifically to test generalization to unknown instructions. Bagel-Robotics-1's scores on the four disclosed suites range from 24.84/19.63% (Memory) to 73.93/58.87% (Precision), all above GPT-6 Astra. Its Open suite score, 4.86/3.67%, is below every other top-tier competitor on the leaderboard.
Read more: Bagel Labs releases WorldDiT, a sub-1B world-action model for robot manipulation
How Fast-WAM and Adaptive Structured Pruning Built This Model
The Fast-WAM architecture was developed by researchers at Tsinghua University's Institute for Interdisciplinary Information Sciences and Galaxea AI. It addresses a practical limitation of WAM-class models: standard WAMs generate future visual predictions at inference time, which adds latency and compute cost. Fast-WAM retains the joint training objective — the model still learns to predict futures during training — but strips out the prediction step at test time. The result is single-pass inference at the cost of discarding a signal the architecture was trained to produce.
Adaptive structured pruning goes further. Rather than training a full-capacity model and then removing individual weights, structured pruning identifies and removes entire functional units — channels, heads, or feed-forward layers — based on estimated importance. The 23.95% parameter reduction Bagel reports is therefore a real reduction in model width, not a sparse approximation of a full-width model.
Recovery training after pruning is standard practice: the pruned model's weights are fine-tuned to compensate for the capacity removed. What recovery training cannot reliably recover is the kind of broad representational coverage that may require the pruned capacity in the first place. This matters for the Open suite specifically, where tasks are designed to probe generalization beyond the training distribution.
Bagel Labs has not published an ablation comparing Bagel-Robotics-1 against a non-pruned baseline of equivalent architecture on the Open suite.
A Leaderboard That Was Already Past GPT-6 Astra Before Bagel Arrived
The RoboDojo leaderboard as of November 18, 2026 reflects a competition that has moved quickly. WALL-OSS-1.0-lite from X Square Robot (Shenzhen) held the number-one position as of early November, and AGIBOT's GO-3 Preview claimed first place across three benchmarks including RoboDojo at its October 29 Innovation Day in Shanghai — though AGIBOT has not published a technical report or independent verification of its scores, and the benchmark entry is marked as a preview. VPP2 from ROBOTERA (a Tsinghua University-affiliated company, also known as Starbot Era), open-sourced on GitHub, previously topped the RoboDojo-Sim leaderboard with an average score of 39.26 and an average success rate of 32.26%.
GPT-6 Astra's RoboDojo results, published in September 2026, show scores across all five suites. Astra's inference latency on RoboDojo is approximately 18.56 seconds per step — a consequence of running a large-scale computer-use model against a robot manipulation benchmark for which it was not specifically trained. The result is that a purpose-built manipulation model substantially smaller than Astra can outperform it on four of five RoboDojo suites, while still trailing on the one suite where the benchmark deliberately withholds task information.
Bagel-Robotics-1 sits at number three on the leaderboard. Its overall average score and success rate place it ahead of GPT-6 Astra but behind the two leaders. The press release framing — "beats GPT-6 Astra on four suites" — is accurate in a narrow sense but omits the competitive context: Astra's RoboDojo performance was already surpassed by multiple models before Bagel-Robotics-1 was released.
Read more: Do World Action Models generalize better than VLAs? A robustness study
WAM-Specific Failure Modes That the Benchmark Does Not Directly Measure
The RoboDojo results for Bagel-Robotics-1 raise two WAM-specific concerns that the benchmark does not directly capture.
The first is the BadWAM attack surface, documented by researchers at the National University of Singapore and Hong Kong Polytechnic University. WAMs derive part of their interpretability appeal from the assumption that a model's imagined future can serve as a check on its intended action. BadWAM demonstrates that this assumption is fragile: small visual perturbations can break the alignment between what a WAM imagines and what it executes, either directly driving the model toward task-failing actions (the action-only attack) or doing so while keeping the imagined future visually plausible (the imagination-preserving attack). The imagination-preserving variant is particularly concerning for deployment because it evades future-based safety checks. These attacks are adversarial scenarios, not standard operating conditions, but they represent a risk category specific to WAM architectures that does not appear in RoboDojo's scoring.
The second is task-incomplete imagination, documented by researchers at Seoul National University and KAIST. This failure mode occurs when a WAM's future prediction is internally consistent and action-consistent — producing no obvious error — but omits the state transition required for task completion. The researchers traced the failure to WAM training adapted for short-chunk control: the model learns to favor locally plausible continuations over task-completing transitions. The Completion Aware Guidance paper describes a training-free sampling method to address this, presented at NeurIPS 2026. Fast-WAM's inference-time removal of future prediction eliminates this specific failure mode from the inference path, but it also removes the future-based verification that might catch other action errors.
The Open Suite Gap and What It Means for Deployment
The five-suite structure of RoboDojo encodes a deliberate hierarchy of difficulty. The four disclosed suites test known task distributions: a model that scores well on Generalization, Precision, Long-Horizon, and Memory has demonstrated competence at tasks within its training scope. The Open suite tests something different — the ability to handle genuinely unknown instructions on first contact.
Bagel-Robotics-1's Open suite score of 4.86/3.67% against GPT-6 Astra's 34.36/31.00% is a large gap on the most deployment-relevant dimension of the benchmark. It suggests that the model's strong scores on the four disclosed suites may reflect effective optimization for known task structures, while Open suite performance — where optimization on known distributions provides less advantage — surfaces a generalization constraint.
Three factors plausibly contribute. First, structured pruning may have removed capacity that contributes disproportionately to generalization; the literature on WAM robustness to perturbations suggests that WAMs' generalization advantage over VLAs depends on the breadth of the visual world-modeling foundation, which pruning directly reduces. Second, Fast-WAM's inference-time omission of future prediction removes a planning mechanism that, in full WAMs, helps the model reason about consequences of actions on novel tasks. Third, the disclosed suites allow submitters to include task-specific data or tune on known distributions before submission; the Open suite specifically blocks this path.
Bagel Labs has not addressed the Open suite result in public statements. The gap between the four-suite performance and the Open suite performance is the article's central finding — and the number the company chose not to highlight.



