Back to News
InnovationAI Understanding briefing

Bagel Labs’ robot policy ranks third on RoboDojo leaderboard, outperforms GPT‑6 Astra in four suites

Bagel Labs’ Bagel‑Robotics‑1 model placed third among learned robot policies on the RoboDojo‑Sim leaderboard, scoring higher than OpenAI’s GPT‑6 Astra in precision, memory, long‑horizon and a custom metric, though it lagged on open‑ended instruction following.

5 min readRead the linked source
Source-provided image accompanying Bagel Labs’ robot policy ranks third on RoboDojo leaderboard, outperforms GPT‑6 Astra in four suites
Source referenceSource recorded
Publisher
techtimes.com
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Key terms

Memory (Agent Memory)
Stored context an AI agent uses across steps or sessions to improve continuity.
Generalization
How well a model performs on new, unseen data outside the training set.
Model Card
Documentation describing a model's intended use, metrics, limitations, and risks.

What happened

Bagel Labs released its Bagel‑Robotics‑1 robot policy on Nov. 19, 2026. The model entered the RoboDojo‑Sim simulation leaderboard on Nov. 18 with an average score of 38.65 and a 31.91% success rate, placing it third among learned policies behind WALL‑OSS‑1.0‑lite and VPP2. In four benchmark suites—Precision, Memory, Long‑Horizon and a custom metric—the model posted higher scores than OpenAI’s GPT‑6 Astra, though Astra achieved a higher success rate on the Open‑suite (31.00% vs. 3.67%). The release includes a technical description of a structured‑pruning pipeline that reduces a 12.47‑billion‑parameter video expert to a 29 GB checkpoint, requiring a 96 GB RTX PRO 6000 Blackwell GPU for inference. The repository is public but does not specify a license for the weights, and real‑robot testing remains unverified.

Bagel Labs announced the Bagel‑Robotics‑1 policy on Nov. 19, 2026, adding it to the RoboDojo‑Sim simulation leaderboard the day before. The model achieved an average score of 38.65 and a 31.91% success rate, placing it third among learned policies, behind WALL‑OSS‑1.0‑lite (42.59) and VPP2 (39.26). The leaderboard distinguishes learned policies from agent systems, which are shown but not ranked.

In four benchmark suites—named Precision, Memory, Long‑Horizon and a custom metric—the model posted higher scores than GPT‑6 Astra (e.g., 34.08% vs. 4.00% in Precision). Astra led only on the Open‑suite, where Bagel‑Robotics‑1 completed 3.67% of episodes versus Astra’s 31.00%. The release does not clarify whether the higher scores reflect success rate or a different metric, but the suite‑by‑suite table suggests Bagel outperforms Astra on the four named dimensions.

Bagel’s technical report describes a three‑step structured‑pruning pipeline: (1) data selection from 30,000 training windows across 35 tasks, (2) adaptive pruning of video‑expert features, feed‑forward units and cross‑attention heads, and (3) recovery training of the compact video expert while keeping the action expert fixed. The resulting video expert contains 12.47 billion parameters, packaged in a 29 GB checkpoint plus 16.7 GB of assets. Inference requires a 96 GB RTX PRO 6000 Blackwell GPU; lower‑capacity GPUs are not supported.

The does not list a license for the released weights, citing upstream terms from components such as VPP2, Fast‑WAM and Wan2.1. The repository is public, but its status statements are contradictory regarding privacy and inference requirements. No real‑robot testing is documented; the scores are based on simulation, and the RoboDojo paper notes that real‑world success rates are typically lower (best reported 12.8%).

Source details: techtimes.com

Why it matters

The announcement shows a North‑American lab competing directly with OpenAI’s flagship GPT‑6 model on a robotics benchmark, suggesting that specialized, structured‑pruned robot policies can achieve competitive performance on precision‑heavy tasks while using fewer parameters. If the model’s claimed hardware requirements hold, it could lower the compute barrier for deploying advanced robot policies in manufacturing or logistics settings that have access to high‑end GPUs. However, the model’s poor performance on open‑vocabulary instruction following highlights a trade‑off between parameter reduction and , a key concern for real‑world robot deployments that must handle unpredictable commands. The lack of a clear weight license and unconfirmed real‑robot testing mean potential adopters must conduct their own due‑diligence, especially in regulated industries where provenance and safety certifications are mandatory. The result also underscores the importance of benchmark transparency: the leaderboard’s average score masks suite‑specific strengths and weaknesses, and the distinction between learned policies and agent systems can affect how rankings are interpreted.

Competing directly with OpenAI’s GPT‑6 Astra on a robotics benchmark demonstrates that specialized, pruned models can achieve comparable or superior performance on certain task families without the massive parameter counts of large language models. This could broaden access to high‑performance robot policies for organizations that can afford a single high‑end GPU but not large‑scale clusters.

The structured‑pruning approach reduces model size while preserving task‑specific accuracy, offering a potential pathway for deploying sophisticated robot policies on edge hardware. However, the model’s weak performance on open‑vocabulary instruction following indicates that pruning may sacrifice flexibility, a critical factor for robots operating in dynamic, unstructured environments.

The unclear licensing and lack of real‑world validation raise compliance and safety concerns. Companies in regulated sectors (e.g., pharmaceuticals, aerospace) will need to verify provenance, conduct latency and reliability testing, and possibly negotiate separate licensing terms before integrating the model into production lines.

The leaderboard’s methodology—averaging scores across five suites and separating learned policies from agent systems—means that a third‑place ranking does not necessarily translate to overall superiority. Stakeholders must examine suite‑specific results to determine suitability for their use cases.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Future updates from Bagel Labs on real‑robot validation, latency measurements on a broader range of GPUs, and any licensing clarification for the model weights. OpenAI may respond with additional GPT‑6 benchmark runs or fine‑tuned variants targeting precision tasks. Industry observers should monitor whether other North‑American robotics firms adopt similar structured‑pruning pipelines to reduce model size without sacrificing task‑specific accuracy, and whether customers in manufacturing or logistics begin pilot projects with Bagel‑Robotics‑1.

Bagel Labs may publish latency benchmarks on a range of GPU configurations, which would clarify the practical deployment cost of the model.

Any licensing clarification or open‑source release of the weight files could affect adoption, especially for academic or commercial partners requiring clear usage rights.

OpenAI could release updated GPT‑6 variants or fine‑tuned models targeting precision and memory tasks, potentially narrowing the performance gap highlighted by Bagel Labs.

Industry uptake: announcements of pilot projects in manufacturing, logistics or warehouse automation that adopt Bagel‑Robotics‑1 would signal market validation and could spur further investment in structured‑pruned robot policies.

Related guides & quizzes

Found this useful?