AnswersPublished17 sources
MagicLab’s EchoPolicy-0 Leads MolmoSpaces; MagicBrain Posts 94.4% on Pick
MagicLab reported EchoPolicy 0 first on AI2’s MolmoSpaces Combined leaderboard at 75.4% across four tasks; its 3.4B parameter MagicBrain ranked fourth at 72.4%, with 94.4% on Pick.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- MagicLab reported EchoPolicy 0 first on AI2’s MolmoSpaces Combined leaderboard at 75.4% across four tasks; its 3.4B parameter MagicBrain ranked fourth at 72.4%, with 94.4% on Pick.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: As reported on October 14, 2026, how did MagicLab’s EchoPolicy-0 and its in-house 3.4-billion-parameter MagicBrain perform on AI2’s MolmoSpa. Article summary: MagicLab reported EchoPolicy-0 first on AI2’s MolmoSpaces Combined leaderboard at a 75.4% equal-weighted success rate, while its 3.4-billion-parameter MagicBrain reportedly ranked fourth at 72.4%, including 94.4% on Pick. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
MagicLab reported that its EchoPolicy-0 robot-manipulation framework ranked first on AI2’s MolmoSpaces Combined leaderboard, with a 75.4% equal-weighted success rate across four tasks. Its 3.4-billion-parameter MagicBrain framework placed fourth at 72.4%, including a reported 94.4% success rate on Pick. The scores are benchmark results—not evidence by themselves of performance in real-world deployment.1
18
How the two systems scored
EchoPolicy-0’s reported average covers MolmoSpaces v1’s Pick, Pick and Place, Open, and Close tasks. MagicLab said it ranked first on the Combined board.1 A review of the published leaderboard data by Pandaily reported a calculated average of about 75.5%, close to the company’s 75.4% figure; the small difference appears in the reporting of the average, so the company-reported figure is retained here.
41
MagicBrain, MagicLab’s 3.4B-parameter system, was reported fourth in the same Combined grouping, with a 72.4% average and 94.4% on Pick.18 That Pick result is a task-specific score; it is not the system’s overall average.
| System | Reported Combined result | Notable detail |
|---|---|---|
| EchoPolicy-0 | 75.4%, first | Equal-weighted average across Pick, Pick and Place, Open, and Close |
| MagicBrain | 72.4%, fourth | 94.4% reported on Pick |
What MolmoSpaces measures
MolmoSpaces is an AI2 zero-shot embodied-AI benchmark built around more than 230,000 indoor environments and about 130,000 object assets.1 It evaluates whether a robot system can carry out instructed manipulation tasks in specified scenes, rather than relying on a single demonstration or task score. Those environment and asset totals describe the benchmark ecosystem, not the number of trials in these results.
The leaderboard figures summarize performance on the benchmark’s four v1 manipulation tasks. They are useful for comparing systems under a shared evaluation, but they do not alone show how reliably a robot would perform in a physical workplace or other deployment setting.
EchoPolicy-0: coordinate understanding and action
MagicLab describes EchoPolicy-0 as a unified framework combining a vision-language model, geometric manipulation tools, and a general action policy. Its process links observing a scene and understanding an instruction with choosing an action, executing it, and checking the result.18
The reported approach also uses visual guidance to choose among candidate actions: an action model proposes options, and the system uses its understanding of the target to select a suitable one. Evaluation documentation describes visual checks that can inform whether execution should continue, be retried, or fall back to another route.18
20 This makes the framework more than a one-shot movement command: it connects perception and action with a feedback step.
MagicBrain: improve target selection, not just movement
MagicBrain takes a different route, using in-context visual guidance to help identify the object relevant to an instruction. In this approach, visual identification guides the action model’s movement, addressing the distinction between recognizing what to handle and deciding how to move.19
33
MagicLab’s reported analysis identified incorrect target selection—not unstable movement—as a common source of grasp failures in the cases it studied.30 That finding points to an instruction-grounding problem: even a well-executed grasp fails if the robot chooses the wrong object. It is a reported observation about those cases, not a universal breakdown of robot-grasp failures.
What the results do—and don’t—show
Both systems reflect a focus on connecting instruction understanding, visual target selection, and physical action. MagicLab Studio has described layered architectures as a preferred direction and instruction following as a challenge for general-purpose tasks.33
The scores show how the systems performed on MolmoSpaces’ benchmark tasks. They should not be read as proof of real-world reliability or commercial readiness: performance in the benchmark does not, on its own, establish how a system will transfer to different physical conditions or deployment settings.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "MagicLab’s EchoPolicy-0 Leads MolmoSpaces; MagicBrain Posts 94.4% on Pick"?
MagicLab reported EchoPolicy 0 first on AI2’s MolmoSpaces Combined leaderboard at 75.4% across four tasks; its 3.4B parameter MagicBrain ranked fourth at 72.4%, with 94.4% on Pick.