MiniCPM-Robot opens a compact embodied AI stack for robots

aigclink · x · 2026-07-20

MiniCPM-Robot is a newly open-sourced robot “brain” from OpenBMB / 面壁智能, with two models:

MiniCPM-RobotManip: a 1.5B VLA model for general manipulation that is said to work across downstream tasks in both simulation and real robots.
MiniCPM-RobotTrack: a 0.9B embodied target-tracking model designed to run fully on-device.

The post highlights several technical points:

RobotManip reportedly scores 91.6 on RoboTwin2 (hard) and 53.3 on RMBench, beating much larger baselines in those benchmarks.
RobotTrack is presented as open-source SOTA on EVT-Bench and can run on the onboard compute of a Unitree Go2 EDU, using only vision and natural-language instructions.
Performance claims include 5+ FPS and about 180 ms end-to-end latency on the robot.
For RobotManip, the key engineering trick is streaming inference: instead of recomputing a long history every step, it incrementally updates context, cutting the cost for 60 frames from 125 TFLOPs to 3.3 TFLOPs.
It also inherits visual token compression from MiniCPM-V 4.6, shrinking per-frame tokens from 256 to 64.
RobotTrack uses a DAgger data loop to improve robustness on long-tail cases such as target crossings, sharp turns, occlusions, and multi-person interference.

The post also mentions HF weights, one-command demo scripts, and a day-one adapter from PhyAI that reportedly pushed inference on H20 from 10Hz to 37Hz.

Original post →

MiniCPM-Robot opens a compact embodied AI stack for robots

More from Embodied