Magic Atom's embodied Agent framework tops MolmoSpaces robot-manipulation leaderboard

Chinese robotics company Magic Atom says its VLM Agent framework EchoPolicy-0 has taken first place worldwide on MolmoSpaces, the embodied-AI zero-shot benchmark run by the Allen Institute for AI (AI2). According to the outlet QbitAI, EchoPolicy-0, built by the company's MagicLab Studio, posted a 75.4% average robot-manipulation success rate across all four v1 tasks in the Combined leaderboard.

On the same leaderboard and in the same group, Magic Atom's in-house embodied Agent framework MagicBrain placed fourth globally with a 72.4% average success rate using only 3.4 billion parameters. Its pick success rate reached 94.4%, which the company says is far above the industry average.

MolmoSpaces is described as one of the most credible zero-shot benchmarks in embodied AI, covering manipulation, navigation and longer-horizon tasks. Its ecosystem spans more than 230,000 indoor environments and roughly 130,000 object assets, forcing robots to find targets across varied backgrounds, geometry and object distributions rather than replaying a fixed demo scene.

The benchmark also includes movable rigid bodies and articulated objects, so a robot must not only identify the right target in an image but also change the object's state through real simulated contact. That makes the tasks genuinely difficult and a stronger test of generalisation.

EchoPolicy-0 completed a full 3,915-episode evaluation across the four MolmoSpaces manipulation tasks. Its core mechanism is VLM goal-guided multi-candidate trajectory selection: the action model proposes several possible motions, projects them onto the image plane, then picks the trajectory that best matches the goal specified by the VLM and keeps verifying after execution. The current version has been validated on four manipulation categories on a Franka single-arm platform.

MagicBrain follows a route the team calls In-Context Visual Guidance, in which the VLA model only handles motion while the job of finding the target goes to a VLM Agent that is better at reading images. It took first place in the Pick-v1 and Pick & Place-v1 task leaderboards, with a pick success rate of 94.4%.

A MagicLab Studio lead said the team favours a hierarchical architecture for the embodied brain and sees instruction following as a key difficulty for current models on general tasks, with EchoPolicy enforcing it through the Agent framework and trajectory scoring and MagicBrain through end-to-end visual guidance.

At industry level, the jump in pick success from roughly 70% to 94.4% means VLA systems are edging from demo-ready prototypes toward production-ready deployment. Picking and placing underpin high-frequency work in factory sorting, warehouse handling and home organising, so every gain in success rate widens the range of commercially viable scenarios.

Magic Atom argues embodied AI should advance through hierarchical collaboration, with upper-level models handling understanding, planning and error correction and lower-level models turning intent into reliable physical motion. The team adds that the visual-guidance technique could migrate to interaction and other scenes, suggesting that a combine-rather-than-one-model-does-all division of labour may be closer to reality than a single generalist model.