MagicLab announced that its agent system based on large vision-language models (VLMs) for general robot control, named EchoPolicy-0, has won first place in the overall MolmoSpaces ranking, which is curated by the Allen Institute for AI (AI2). The average success rate was 75.4% across all four manipulation tasks in version v1.
MolmoSpaces is a benchmark for embodied AI, created based on over 230,000 internal environments and approximately 130,000 objects. Pandaily verified the stated data by cross-referencing it with information published by AI2 for its ranking.
EchoPolicy-0 is registered under MagicLab and utilizes the Gemini-3.8-Flash model as its vision-language model. It was tested on the Franka manipulator across all four tasks: Pick, Pick and Place, and Open and Close. Averaging the success rates across these four tasks yields approximately 75.5%, which aligns with the company's stated 75.4% after rounding, and is the highest among all presented works covering all four tasks.
A key feature of EchoPolicy-0 is the integration of a vision-language model, geometric manipulation tools, and a general action policy into a single closed loop that includes observation, goal understanding, skill selection, action execution, and result verification. If an action proves ineffective, the framework is capable of continuing operation, retrying, replanning, or rolling back. The core concept of the system is target trajectory selection: the action model proposes several possible movements, which are then projected onto the image, and the vision-language model selects the most suitable one for the goal before execution verification takes place.
Furthermore, MagicLab introduced MagicBrain—its proprietary system with 3.4 billion parameters, which the company claims ranks fourth overall. AI2 data shows that MagicBrain achieves a 94.4% success rate in the Pick task. The approach, which MagicLab calls contextual visual guidance, allows the VLM agent to locate the target in the image while the action model focuses exclusively on movement. The company notes that most failed grasps they studied occurred due to selecting the wrong object, rather than motion instability.
The head of the MagicLab Studio research lab stated that the team prefers a multi-level architecture and considers following instructions one of the main challenges for existing models. EchoPolicy-0 addresses this problem through agent orchestration and trajectory evaluation, while MagicBrain uses end-to-end visual guidance. It should be noted that the results obtained are from simulation tests, and MagicLab has not provided information regarding the timeline for integrating these frameworks into its commercial robots.