Magic Atom said on Oct. 19 that it has launched the MagicArm R1, a flexible biomimetic force-controlled robotic arm, and built a desktop robotic-arm solution around a vision-language-action (VLA) embodied AI model.
The solution uses the MagicArm R1 as its physical platform and runs on EchoPolicy-0, the company’s embodied AI model, which ranked first on the MolmoSpaces Combined leaderboard with a 75.4% average success rate in robot manipulation. MolmoSpaces is a global zero-shot embodied AI benchmark initiated by the Allen Institute for AI (AI2).
Magic Atom said the system integrates visual perception, intelligent decision-making, motion control, data collection, model training and real-robot reinforcement learning into a single hardware-software platform. It is aimed at robot algorithm research, model training, intelligent operation validation and industrial technology pre-research.
The MagicArm R1 has six degrees of freedom, a 700 mm reach and weighs 5.5 kg, with a 4 kg payload. Its forearm can bend up to 180 degrees. The company said the arm has repeat positioning accuracy of ±0.1 mm and can suppress end-effector vibration within 0.4 seconds.
Unlike traditional desktop robotic arms that rely mainly on rigid position control, the R1 introduces real-time force sensing at all joints. Magic Atom said this is intended to address a long-standing trade-off between precision and compliance in tasks such as grasping fragile objects, insertion assembly and drag teaching.
The arm supports upright, side-mounted and inverted installation, and is compatible with grippers, dexterous hands and suction cups. It also has built-in multi-tool management logic, the company said.
On the software side, MagicArm R1 offers an open SDK and an AI development toolchain covering data collection, model training and deployment evaluation. It provides URDF files and supports real-robot reinforcement learning and DAgger closed-loop training, with force and pose data output in real time for model development.
Magic Atom said EchoPolicy-0 combines vision-language model (VLM) understanding with VLA decision-making. The VLM can interpret instructions, identify target objects and analyze spatial relationships, and participates in action evaluation through a “VLM-guided multi-candidate trajectory selection” method. The system generates multiple possible motions, projects them onto the image plane, and uses the VLM to select trajectories that better match the task target.
The desktop solution also uses a Progress task-progress model to automatically assess and filter teleoperation data, excluding low-quality segments such as pauses, jitter and repeated adjustments. Combined with real-robot reinforcement learning and a “deployment-collection-training-evaluation-redeployment” iteration mechanism, the company said the robot can accumulate real interaction data and improve its capabilities through use.
MagicLab Studio, Magic Atom’s R&D laboratory, said the desktop robotic-arm solution can be used for robot algorithm research, model training, intelligent operation validation and industrial technology pre-research. In industrial pre-research, it can support technical validation for tasks including material sorting, small-part transfer and precision assembly, as well as drag teaching for rapid deployment and human-robot collaborative workstations.
Magic Atom said the R1 also supports secondary development and can be mounted on mobile platforms such as quadruped robots. The company said the integrated desktop format lowers the threshold for embodied AI development, experimentation and deployment compared with building hardware systems, data-collection tools and training environments separately.
AGIBOT
Figure AI
UBTECH