SSyncDev

MiniCPM-RobotTrack

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

MiniCPM-RobotTrack is OpenBMB's compact vision-language-action policy on MiniCPM4-0.5B that follows a described person on a Unitree Go2 at 5+ FPS on board.

Key specifications

Specification score
9.3
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights under Apache 2.0; no hosted price.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

VisionOpen weights

What is MiniCPM-RobotTrack?

"Follow the person in the red shirt." That one sentence is the whole interface to MiniCPM-RobotTrack. OpenBMB's compact vision-language-action model reads a natural-language instruction and a stream of camera features, and returns eight future waypoints, each an x, y and yaw, for a legged robot to walk. It runs fully on the robot. OpenBMB reports a stable 5+ frames per second with about 180 ms end-to-end latency on the Unitree Go2's onboard compute.

It launched on 18-19 July 2026 as one half of MiniCPM-Robot, alongside the manipulation model MiniCPM-RobotManip.

Size: read the fine print

The language backbone is MiniCPM4-0.5B, and the stored checkpoint holds about 450M parameters. OpenBMB's own comparison table lists the model as 0.9B. The difference is most likely the visual feature extractors (DINOv3 and SigLIP), which run in the preprocessing pipeline rather than inside this checkpoint, but OpenBMB does not spell it out. Budget memory for both the policy and the encoders.

Inputs and outputs

Repository openbmb/MiniCPM-RobotTrack
Instruction natural language
Vision features fused DINOv3 + SigLIP, 1,536-dimensional
History 31 past frames at 4 coarse tokens each
Current frame 64 fine tokens
Output 8 waypoints of [x, y, yaw]
Licence Apache 2.0

Publisher-reported EVT-Bench results

Success rate, tracking rate and collision rate (lower is better), in percent:

  • Single-target tracking: 84.1 / 89.8 / 3.0
  • Distracted tracking: 53.2 / 73.4 / 13.6
  • Ambiguity tracking: 58.0 / 80.4 / 9.0

Among the open checkpoints OpenBMB compares, it leads on single-target tracking and on distracted-tracking success and tracking rate. Closed TrackVLA variants still beat it on success in several splits. The distracted setting's 13.6% collision rate is the number to watch if your environment is crowded.

Why the data loop matters

OpenBMB trained it with a DAgger-style cycle: let the model act in simulation and on real robots, find failures such as target crossings, sharp turns, brief occlusions and several people crossing paths, correct them with expert policies or rules, and fold them back into training. Automated checks and manual review filtered bad trajectories. The demos cover outdoor obstacles, elevators and underground car parks.

Before you put it on a robot

Test in your own corridors with deliberate occlusions and two similarly dressed people, and log collisions as well as losses of track. The deployment scripts target the Go2 EDU specifically; other platforms will need integration work. For arm manipulation, see MiniCPM-RobotManip. Other OpenBMB models are on the OpenBMB provider page, and related policies are under reinforcement learning.

How to evaluate MiniCPM-RobotTrack

Compare the model on the work you need to complete.

Tasks and inputs

MiniCPM-RobotTrack is listed for Reinforcement learning, Video classification. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Eight future [x, y, yaw] waypoints for following a target described in natural language.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

runway

Praxis-1

Praxis-1 is Runway's open 27B world action model for robotics, predicting video and robot actions together; Apache 2.0, LeRobot and vLLM-Omni support.

— context

NVIDIA

Metatron-v1

Metatron-v1 is NVIDIA's open unified humanoid policy for the Unitree G1 that walks, interacts with scenes and recovers from falls, even in a browser.

— context

sensetime

SenseNova-RoboRSI-27B-Preview

SenseNova-RoboRSI-27B-Preview is SenseTime's open 27B vision-language model for embodied agents, post-trained from Qwen3.8-27B by recursive self-improvement.

— context

OpenBMB

MiniCPM-RobotManip

MiniCPM-RobotManip is OpenBMB's open 1.5B vision-language-action model for robot manipulation, keeping a minute of visual memory on-device.

— context

DocPereira

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context