SSyncDev

MiniCPM-RobotManip

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

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

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: No API pricing stated; open weights for self-hosting

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

  • Image input
  • Text input

Produces

  • Text output

Capabilities

VisionOpen weights

What is MiniCPM-RobotManip?

Most robot policies decide from the current camera frame and forget everything before it. That is fine for picking up a cup and useless for "put back what you took out earlier". MiniCPM-RobotManip keeps a running visual history, about a minute of it, while staying small enough for on-device control.

OpenBMB released the 1.5B vision-language-action model on 18 July 2026 under Apache 2.0, together with a sibling, MiniCPM-RobotTrack.

Three claims worth unpacking

One generalist policy. A single set of weights serves every downstream task. OpenBMB reports it outperforms larger models such as π0.5 and Qwen-VLA on representative evaluations, which are publisher-run.

Streaming context. Past observations are folded into the model's context as it runs. Per-step compute drops from 125 TFLOPs to 3.3 TFLOPs while keeping 60 frames of history, enough for roughly one minute of visual memory. That is the difference between reacting and deciding with context.

Compressed vision. It inherits MiniCPM-V 4.6's token compression, cutting each frame from 256 to 64 visual tokens.

Latency

Measure MiniCPM-RobotManip π0.5
Forward latency per decision step (H100, BF16, single frame) 120 ms 234 ms

The figure excludes autoregressive decoding of the task, so full control-loop latency will be higher. Measure on your own robot computer.

Integration notes

  • Requires transformers==5.7.0 exactly.
  • Loaded with AutoModel and AutoProcessor using remote code.
  • The reference example uses an 80-dimensional state vector and 448×448 images.
  • Your robot's state layout must be mapped to what the checkpoint expects.

How to try it responsibly

Start in simulation or on a table-top arm with speed and force limits. Pick tasks that need memory, such as returning objects to where they started, and compare with a single-frame policy on the same tasks. Track success rate, recovery after a failed grasp and control frequency. Small language models from the same lab, such as MiniCPM5 1B, show the family's on-device focus; more releases are on OpenBMB's provider page.

How to evaluate MiniCPM-RobotManip

Compare the model on the work you need to complete.

Tasks and inputs

MiniCPM-RobotManip is listed for Image-text-to-text, Reinforcement learning. It accepts image, text 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

About 1.5B parameters, a single generalist vision-language-action policy.

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-RobotTrack

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.

— 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