MagicLab says its VLM agent framework for general robot manipulation, EchoPolicy-0, has taken first place on the MolmoSpaces Combined leaderboard run by the Allen Institute for AI (AI2), with an equal-weighted average success rate of 75.4% across all four v1 manipulation tasks. MolmoSpaces is a zero-shot embodied AI benchmark built on more than 230,000 indoor environments and about 130,000 object assets.
Pandaily checked the claim against the data AI2 publishes for its leaderboard. EchoPolicy-0 is listed under MagicLab, paired with Gemini-3.8-Flash as the vision-language model and evaluated on a Franka arm, in all four tasks: Pick, Pick and Place, Open and Close. Averaging the four task success rates in that data gives about 75.5%, consistent with the company's 75.4% figure after rounding, and the highest among entries that cover all four tasks.
EchoPolicy-0 links a vision-language model, geometric manipulation tools and a general action policy into one closed loop of observing, understanding the goal, choosing a skill, acting and checking the result. When an action falls short, the framework can continue, retry, replan or roll back. Its key idea is target-guided trajectory selection: the action model proposes several candidate motions, these are projected onto the image, and the vision-language model picks the one that best matches the target before execution is verified.
MagicLab also entered MagicBrain, a 3.4-billion-parameter system it developed in-house, which the company says ranks fourth on the combined board. AI2's data lists MagicBrain with a 94.4% success rate on the Pick task. Its approach, which MagicLab calls in-context visual guidance, lets a VLM agent locate the target in the image while the action model focuses only on movement. The company says most grasp failures it studied came from picking the wrong object rather than unstable motion.
The head of MagicLab Studio, the company's research lab, said the team favors a layered architecture and sees instruction following as a key difficulty for current models. EchoPolicy-0 addresses it through agent orchestration and trajectory scoring, and MagicBrain through end-to-end visual guidance.
The results are simulation benchmarks. MagicLab has not said when either framework will be deployed on its commercial robots.
SEE ALSO: Tsinghua AIR Zeva Lifts Embodied Manipulation Success From 26% to 73% Without Retraining · MagicLab Demonstrates Large-Scale Robot Coordination in Outdoor Performance