MiniCPM-Robot Brings a 1.5B VLA and 0.9B Tracker to On-Device Robotics

ModelBest and OpenBMB released MiniCPM-Robot on July 19, their first open model family built to perceive, decide, and act in the physical world. The release combines a 1.5-billion-parameter manipulation model with a 0.9-billion-parameter target tracker.

That pairing creates a more interesting contest than another benchmark race. MiniCPM-Robot tests whether small, open models can deliver useful robotic behavior without depending on large servers, proprietary access programs, or continuous network connections.

The family enters a field led by ambitious systems such as Physical Intelligence's π0.5, Qwen-VLA, and Google DeepMind's Gemini Robotics models. Those projects pursue broad generalization across tasks and machines, but they differ sharply in size, availability, and deployment strategy.

MiniCPM-Robot takes a narrower route. It separates manipulation and tracking, then concentrates on compact models, local execution, memory efficiency, and reproducible deployment. That approach lowers one barrier, but it does not settle the harder question of reliability outside selected tests.

MiniCPM-Robot Splits Physical Intelligence Into Two Small Models

The central change is not simply that ModelBest entered robotics. It released weights and implementation code for two distinct physical tasks.

The first model, MiniCPM-RobotManip, is a 1.5-billion-parameter vision-language-action model. A VLA converts visual observations and language instructions into actions that a robot can execute.

ModelBest describes RobotManip as a generalist policy, meaning one set of weights supports multiple manipulation tasks. Its published examples include object handling and longer sequences, such as preparing a sandwich.

RobotManip builds on MiniCPM-V 4.6, according to the company. It extends the earlier multimodal system from interpreting images and language into predicting physical actions.

The second release, MiniCPM-RobotTrack, focuses on mobile target tracking. It interprets a natural-language instruction, identifies the intended person, and produces movement commands while responding to obstacles and visual interference.

RobotTrack uses a MiniCPM4-0.5B language backbone, but the complete policy contains 0.9 billion parameters. It combines that backbone with visual components and a control pipeline for navigation.

The distinction matters because manipulation and tracking place different demands on a model. A robot arm needs precise object interactions, while a mobile tracker must update continuously as people and obstacles move.

Rather than claiming one universal policy already handles both jobs, ModelBest released specialized systems under one family. This makes the project easier to inspect, although it also limits claims about unified physical intelligence.

The official MiniCPM-Robot repository includes checkpoints, inference code, training entry points, evaluation scripts, and deployment guidance. The code and model weights carry an Apache 2.0 license.

The repository also documents a supporting inference framework called PhyAI. ModelBest says its optimizations increased RobotManip throughput on an Nvidia H20 from 10 hertz to 37 hertz.

That number applies to a specified software and hardware configuration. It should not be read as a general speed guarantee across different robots, GPUs, or control systems.

The release follows ModelBest's broader focus on smaller models for local use. MiniCPM-Robot applies that same design preference to systems where latency, network availability, and onboard compute directly affect behavior.

This is ModelBest's first public embodied model family, according to its announcement. The company reportedly formed its dedicated embodied intelligence team in January 2026, only months before the release.

That short development period is notable, but it also raises the standard for outside validation. Public code permits scrutiny, while sustained real-world testing will determine whether the results transfer beyond the demonstrated platforms.

The 1.5B VLA Makes Memory an Efficiency Problem

MiniCPM-RobotManip's main argument is that a robot can retain useful visual history without repeatedly processing every previous frame.

Many robot policies react primarily to the latest observation. That design can work for simple movements, but long tasks require remembering completed steps, changing object positions, and earlier instructions.

RobotManip places historical observations into a streaming context. Streaming inference reuses previously computed information instead of reconstructing the entire visual history for every new decision.

ModelBest says processing 60 historical frames through conventional recomputation would require 125 trillion floating-point operations per decision. Its streaming method reportedly reduces that figure to 3.3 trillion operations.

The company also claims that RobotManip can retain up to one minute of visual context. Online processing costs remain comparable to a reactive policy using one current frame, according to its documentation.

Those claims define the model's most consequential mechanism. A smaller parameter count matters less if a robot repeatedly spends enormous compute on its expanding visual history.

The system also compresses each frame from 256 visual tokens to 64. Tokens are compact units the model processes after encoding an image, so fewer tokens generally reduce computation.

On an Nvidia H100 using BF16 precision and one input frame, ModelBest reports 120 milliseconds of model-forward latency per decision. The comparable figure listed for π0.5 is 234 milliseconds.

The measurement excludes autoregressive action decoding. It therefore describes only part of the complete control loop, not the time between a camera observation and physical motion.

RobotManip returns a chunk containing 30 steps with 80 action dimensions in the provided inference example. Action chunking lets a system predict a short sequence instead of generating one movement at a time.

The benchmark results are similarly specific. RobotManip scores 97.5 on LIBERO and 4.1 on the Calvin ABC-to-D evaluation, according to the repository.

On RoboTwin2, it records 91.3 on the easy setting and 91.6 on the hard setting. These figures place it near or above several larger systems within the listed evaluation configurations.

Its largest reported advantage appears on RMBench, which tests contextual memory for robot manipulation. RobotManip scores 53.3, compared with 10.4 for π0.5 in ModelBest's published comparison.

Physical Intelligence designed π0.5 for open-world generalization. Its research combines robot data, web information, semantic subtask prediction, and heterogeneous training across multiple embodiments.

That broader design makes π0.5 an important reference, but the two systems should not be reduced to one score. Their training mixtures, evaluation settings, architectures, and intended deployment environments differ.

RobotManip is also smaller than the 5B-plus Qwen-VLA system listed in ModelBest's comparison. The Qwen-VLA research unifies manipulation, navigation, and trajectory prediction through embodiment-aware conditioning.

Qwen reports results across real robots, simulators, navigation tasks, and changing environments. MiniCPM-RobotManip instead emphasizes compact manipulation, visual history, and openly downloadable weights.

The real contest is therefore not 1.5 billion parameters against five billion. It is focused efficiency against a broader attempt to unify physical tasks and robot forms.

RobotManip's results support the efficiency case within the disclosed benchmarks. They do not establish that the smaller model matches larger competitors across every embodiment, environment, or long-horizon task.

MiniCPM-RobotTrack Moves the Contest Onto the Robot

RobotTrack is the more concrete deployment story because it runs a complete vision-only tracking loop on a commercially available quadruped.

ModelBest validated the model on a Unitree Go2 EDU equipped with a Jetson Orin NX containing 16GB of memory. The documented setup also uses an Intel RealSense D435i camera.

The model processes camera input, recognizes a language-specified target, predicts a waypoint, and sends rate-limited movement commands. ModelBest reports stable operation above five frames per second.

Its disclosed end-to-end latency is approximately 180 milliseconds. Unlike the partial RobotManip latency measurement, this figure covers the wider perception and control pipeline described in the deployment documentation.

The system operates locally and does not require a cloud connection. That makes the design relevant in parking structures, elevators, industrial buildings, and outdoor areas where connectivity can disappear.

Local operation also reduces the delay and uncertainty associated with sending images to a remote service. It does not automatically make the system private, secure, or safe.

RobotTrack uses a data process based on DAgger, short for Dataset Aggregation. The technique lets a policy act in an environment, exposes its errors, and adds corrected examples to later training rounds.

ModelBest says its pipeline targets difficult events such as rapid turns, short occlusions, crossing targets, and crowded intersections. Automated checks and manual review remove invalid interactions and abnormal trajectories.

This matters because ordinary demonstrations often underrepresent failures. A tracking model may look competent until two people cross paths or the chosen target disappears briefly behind an obstacle.

The company evaluates RobotTrack on EVT-Bench across three categories. These cover standard target tracking, interference from distractors, and ambiguous target conditions.

Results use success rate, tracking rate, and collision rate. Higher success and tracking rates are better, while a lower collision rate is preferable.

RobotTrack posts an 89.8 percent tracking rate on standard tracking. It records 73.4 percent during distractor tracking and 80.4 percent under ambiguous conditions.

Among the open-weight systems in ModelBest's comparison, RobotTrack leads standard tracking across all three reported metrics. It also leads the listed open models on distractor success and tracking rates.

The picture becomes more mixed outside those selected comparisons. Qwen-RobotNav records a 90.0 percent standard tracking rate, slightly above RobotTrack, but its weights are not open.

RobotTrack also posts a 13.6 percent collision rate during distractor tracking. That exceeds OmTrackVLA's listed 11.3 percent and deserves attention in any physical deployment.

Under ambiguous conditions, RobotTrack records a 9.0 percent collision rate. OmTrackVLA reports 7.6 percent in the same published table.

These results illustrate the tradeoff clearly. RobotTrack combines a small footprint, open weights, and local execution, but it does not dominate every safety-relevant metric.

The repository defaults its real-robot setup to dry-run mode, which prevents movement commands from reaching the robot. Documentation tells operators to validate latency, cameras, and emergency stopping before enabling live control.

That caution is essential. A visually convincing follow demonstration does not prove safe operation around children, workers, traffic, stairs, or unfamiliar machinery.

Open Weights Pressure Private Robotics Access

MiniCPM-Robot pressures competitors through inspectability and deployment access, not through a proven claim of universal robotic competence.

Google DeepMind has also made local robotics a major theme. Gemini Robotics On-Device follows language instructions and performs dexterous manipulation without relying on a network connection.

Google says developers can adapt its on-device model to new tasks with 50 to 100 demonstrations. The company has shown tasks such as folding clothing, unzipping bags, and pouring dressing.

However, access began through a trusted tester program and SDK. ModelBest provides downloadable checkpoints and code, giving independent developers a different starting point.

Open access lets a laboratory inspect inputs, modify control logic, profile latency, reproduce evaluations, and test failure cases. It also helps smaller robotics teams avoid dependence on a remote model provider.

That freedom does not remove infrastructure requirements. RobotManip's documented environment expects Python 3.10, PyTorch 2.6, and CUDA 12.4.

RobotTrack evaluation requires simulator assets and additional vision models, including DINOv3 and SigLIP. Its Go2 deployment involves TensorRT, ROS 2, Jetson software, cameras, and hardware-specific configuration.

The project is open, but it is not plug-and-play for an inexperienced user. Robotics still demands integration work that ordinary language-model demos largely conceal.

Open weights also shift responsibility. A developer can alter a model or deploy it in an unsupported environment without the original team reviewing the resulting behavior.

This makes documentation, licensing, safety constraints, and reproducible tests especially important. The repository's dry-run default is a sensible starting point, not a complete safety system.

For established robotics labs, MiniCPM-Robot offers another baseline to examine. For smaller teams, its value may lie in shortening the route from downloaded weights to an instrumented physical test.

The release also signals a wider movement toward local robot inference. Cloud models offer abundant compute, but network delay and outages conflict with continuous physical control.

A recent embodied runtime paper describes this deployment problem as a fragmented collection of model-specific Python stacks, backends, and robot integrations. Physical control also demands latency-first, batch-one execution.

MiniCPM-Robot does not solve that industry-wide fragmentation. Its separate environments, dependencies, and deployment paths actually show how much integration remains.

Still, the project moves more of that work into public view. Researchers can inspect not only a checkpoint, but also the surrounding inference, data, evaluation, and deployment choices.

That transparency is the release's strongest competitive pressure. Private systems now face a clearer question about which benefits justify restricted access or undisclosed implementation details.

The Published Benchmarks Leave Major Questions Open

The release is credible enough to test, but its evidence does not yet support broad conclusions about dependable general-purpose robots.

Most headline results come from ModelBest's own repository and announcement. They have not been independently reproduced across multiple laboratories, robot bodies, and uncontrolled environments.

Benchmark scores can also hide substantial differences in evaluation settings. A model labeled generalist may use one checkpoint, while a specialist receives training designed for a narrower task.

Hardware measurements require similar care. The RobotManip latency comparison uses an H100 and excludes action decoding, while RobotTrack's end-to-end result comes from a Jetson-based quadruped stack.

Neither figure describes performance on a typical laptop. They also do not establish energy use, heat constraints, sustained operation, or recovery after a software fault.

RobotManip's strongest result, its 53.3 RMBench score, supports the claim that streaming history helps. Yet the absolute number also leaves considerable room for improvement.

A robot working through a long household or industrial procedure must recover from interruptions, misplaced objects, failed grasps, and human intervention. A memory benchmark covers only part of that problem.

The sandwich demonstration gives the release a tangible scenario. It does not establish a completion rate across kitchens, object arrangements, lighting conditions, tools, or unexpected changes.

RobotTrack presents a similar gap. Following a selected person in an elevator or parking area shows resilience to weak connectivity, but deployment safety requires far more than target retention.

Collision rates remain nonzero in all three published RobotTrack categories. The highest reported rate is 13.6 percent in the distractor scenario.

The practical meaning depends on the benchmark's collision definition and episode design. Operators should not translate that percentage directly into expected incidents in a real building.

The company also calls RobotTrack the first fully local, vision-only, natural-language tracking model of its kind. That positioning depends on category definitions and should be treated as a company claim.

No public release can eliminate misuse. A local tracker could support inspections and assistance, but similar capabilities raise concerns involving surveillance, consent, and operation around unaware bystanders.

Visual tracking also inherits weaknesses from its cameras and encoders. Poor lighting, reflective surfaces, similar clothing, occlusion, and camera obstruction can affect target identification.

The deployment guide names an emergency stop and an onsite operator. Those controls reveal an important truth about present robotics: model intelligence does not replace operational safeguards.

RobotManip introduces another unresolved issue through memory. Retaining visual context can improve task continuity, but longer context may also preserve an early misunderstanding.

If the system mistakes an object or instruction, later actions can build on that error. Efficient memory therefore needs evaluation alongside correction and forgetting behavior.

The Chinese launch coverage reports that the company formed its embodied team in January. Fast execution is encouraging, but maturity requires repeated testing and field feedback.

The proper reading is neither dismissal nor acceptance. MiniCPM-Robot offers enough artifacts for developers to test its claims, and the next evidence should come from those outside tests.

Three Signals Will Show Whether MiniCPM-Robot Matters

The next stage depends on reproducibility, hardware coverage, and evidence from sustained physical deployments.

The first signal is independent benchmark reproduction. Researchers should rerun RobotManip and RobotTrack using the disclosed checkpoints, scripts, and evaluation settings.

Matching the published numbers would strengthen ModelBest's efficiency argument. Large deviations would point toward hidden assumptions, fragile dependencies, or incomplete reporting.

Independent tests should include more than final averages. Per-task outcomes, failed episodes, latency distributions, collision types, and recovery behavior would reveal where the models break.

The second signal is support beyond the initial hardware and benchmark environments. RobotTrack currently has its clearest deployment path on the Unitree Go2 EDU and Jetson Orin NX.

A broader release should demonstrate predictable behavior across different cameras, compute modules, mobile bases, and environmental conditions. RobotManip needs similar evidence across more arms and control formats.

Successful ports would show that the compact models offer transferable deployment value. Difficult ports would suggest that much of the result depends on system tuning around specific hardware.

The third signal is whether open development produces useful improvements. That includes community training recipes, verified checkpoints, safety tests, integrations, and documented fixes for failure cases.

Repository activity alone is not enough. Stars and forks measure attention, while merged contributions and reproducible deployments reveal whether a project is becoming shared infrastructure.

Competitor responses will provide another useful reference inside these three signals. Qwen-VLA pursues a larger unified model, while π0.5 focuses on open-world generalization across complex household tasks.

Google's on-device program shows that local inference is not a niche preference. The competitive question is which route combines access, generalization, latency, and physical reliability.

MiniCPM-Robot currently has the clearest case on access and compactness. It has promising company-reported evidence on memory efficiency, tracking, and selected manipulation benchmarks.

Its weakest case is broad independent validation. No parameter count or open license can substitute for repeated success on unfamiliar robots in changing environments.

Developers evaluating the project should begin in simulation or dry-run mode. They should record latency, target loss, collisions, action errors, and recovery attempts before enabling motion.

Teams should also separate model-forward performance from complete system performance. Cameras, encoders, networking, control limits, actuators, and emergency systems all shape real behavior.

The release matters because it turns a robotics claim into something outsiders can inspect. Its lasting importance will depend on what those outsiders can reproduce, extend, and safely deploy.

MiniCPM-Robot's 1.5B VLA and 0.9B tracker make small open models credible participants in embodied AI. They do not yet make compact general-purpose robots a settled engineering problem.

The useful question now is practical: can independent teams reproduce these results, expose the failure cases, and make the same models work safely on different machines?