WAIC 2026 | ModelBest unveils its first embodied AI achievement: the MiniCPM-Robot series is officially released
On the morning of July 19, the "Intelligence at the Edge, Benefit for Every Industry" sub-forum hosted by ModelBest concluded successfully at the World Artificial Intelligence Conference. At the forum, ModelBest and OpenBMB officially released and open-sourced MiniCPM-Robot, its first embodied AI technical achievement — the "Little Steel Cannon" has officially entered the robotics field!
The MiniCPM-Robot series released this time includes two embodied AI models: MiniCPM-RobotManip, a general-purpose VLA model, and MiniCPM-RobotTrack, aimed at tracking and navigation for mobile robots.
MiniCPM-RobotManip was trained on MiniCPM-V 4.6, ModelBest's latest multimodal on-device model, inheriting the "small size, high performance" and extreme visual token compression advantages of MiniCPM-V 4.6. With only 1.5B parameters, it achieves performance comparable to larger 3-4B models while halving the compute cost of streaming real-time inference. The model supports one minute of embodied native memory, enabling it to efficiently complete manipulation tasks that test contextual memory.
This article includes video content. Please view it in the original WeChat post.
MiniCPM-RobotManip making a sandwich on a real robot (5× speed)
MiniCPM-RobotTrack was developed jointly by ModelBest and Nanjing University. It is the industry's first tracking model that supports real local offline deployment, achieving purely visual, local autonomous instruction tracking for the first time. Natively adapted to the Unitree Go2 Edu, it can perform single-target, dynamic multi-target, and ambiguous-target tracking using only the stock camera and local compute, stably reaching 5+ Hz with an end-to-end response latency of about 180 ms. It also comes with a self-evolving data pipeline that continuously improves tracking in complex dynamic environments.
A robot dog equipped with MiniCPM-RobotTrack can follow instructions zero-shot in open environments, resist interference from people crossing its path, and follow smoothly in no-network or weak-network environments such as elevators and parking garages.
The MiniCPM-Robot series models are all open-sourced. Developers and embodied AI practitioners are welcome to download, deploy, and try them:
➤ GitHub
🔗 https://github.com/OpenBMB/MiniCPM-Robot
➤ Hugging Face
🔗 https://huggingface.co/openbmb/MiniCPM-RobotManip
🔗 https://huggingface.co/openbmb/MiniCPM-RobotTrack
General-purpose VLA: MiniCPM-RobotManip
MiniCPM-RobotManip is a 1.5B general-purpose VLA model for robotic manipulation. It retains the general understanding capability of ModelBest's multimodal foundation model MiniCPM-V 4.6 and is trained uniformly across multiple tasks, so a single model learns to handle different instructions and manipulation tasks.
According to mainstream VLA benchmarks, MiniCPM-RobotManip ranks in the first tier of VLA models in overall performance, close to or even better than mainstream models such as Qwen-VLA and π0.5.
Especially on RMBench, which tests the contextual memory of VLA models, MiniCPM-RobotManip clearly outperforms several mainstream models — it scores 53, while the mainstream π0.5 scores only 10, close to random. As shown below, the model can first cover blocks of different colors and then uncover them in red-green-blue order, demonstrating its native contextual capability, which current mainstream single-frame reactive VLAs cannot do.
MiniCPM-RobotManip's native memory capability (8× speed)
MiniCPM-RobotManip further applies MiniCPM-V 4.6's efficient visual token compression to robotics, greatly reducing the visual input compute and inference latency of the robot. As a result, the inference cost after retaining one minute of contextual memory is comparable to that of traditional single-frame reactive inference — in other words, better model performance at lower deployment cost.
Tests show that in a robotic manipulation task with 60 frames of historical observations, the traditional recomputation approach needs 125 TFLOPs per decision step, while streaming inference needs only 3.3 TFLOPs — lower than π0.5's 5.0 TFLOPs with no historical frames as input. On H100 at BF16 precision, MiniCPM-RobotManip's per-decision-step inference latency is 120 ms, nearly half of π0.5's 234 ms, balancing contextual memory and response efficiency.
The first locally deployable offline tracking model: MiniCPM-RobotTrack
MiniCPM-RobotTrack is a 0.9B end-to-end visual instruction tracking model with three technical advantages:
01 Leading open-source SOTA on three tracking tasks
We evaluated MiniCPM-RobotTrack across three typical scenario dimensions: single-target tracking (STT), dynamic multi-target tracking (DT), and ambiguous-target tracking (AT).
With only 0.9B parameters and no downstream reinforcement learning optimization, MiniCPM-RobotTrack achieved tracking rates of 89.8, 73.4, and 80.4 on the three tasks, the best results among open-source solutions, surpassing OmTrackVLA by 7.0, 14.6, and 6.5 percentage points respectively.
Under the same single-view, pure SFT (supervised fine-tuning) setting, compared with another closed-source model TrackVLA++, MiniCPM-RobotTrack improved tracking rates by 8.8 and 17.0 percentage points on single-target tracking and ambiguous-target tracking, with a 58.0% success rate on ambiguous-target tracking. This shows that with high-quality data and end-to-end training, a lightweight model can achieve competitive real-world instruction tracking without multi-view input or downstream RL.
02 A self-evolving data pipeline that makes the model stronger with use
To address the high cost of real-robot data collection and limited coverage of long-tail scenarios, MiniCPM-RobotTrack builds a self-evolving data pipeline. It first cleans low-quality data such as abnormal trajectories and erroneous actions through automatic detection and manual review, then introduces a DAgger strategy tailored to embodied tracking, letting the model actively expose weak spots during simulation and real-robot interaction. It continuously adds high-value samples for complex scenarios such as sharp turns, short occlusions, and multi-person crossings, improving tracking and generalization within a data loop.
03 Real-robot tests at 5+ Hz, autonomous tracking even offline
After systematic optimization of the full Unitree Go2 pipeline, MiniCPM-RobotTrack stably reaches 5+ Hz on the robot dog's native local compute, with an end-to-end response latency of about 180 ms.
In a live demo, even after the network card was unplugged on site, the robot dog continued to identify targets, track them, and control its motion based on local visual frames and text instructions. This capability can be applied to building inspection, companion care, and campus security, and is expected to extend to disaster rescue, special operations, and other scenarios with weak or no network and strong interference.
This open-source release also provides a complete deployment workflow and one-click startup scripts covering model loading, camera access, text instruction input, and robot control interfaces, making purely local deployment truly ready out of the box.
PhyAI: an embodied AI inference framework
Both models released this time are supported by PhyAI for inference.
PhyAI is an open-source inference framework for Physical AI, designed for extremely fast on-device inference and large-scale RL rollout in the cloud. Compared with the official model repositories, PhyAI running π0, π0.5, and GR00T on an NVIDIA GeForce RTX 5090 achieves speedups of about 2.85×, 1.82×, and 2.28× respectively; GR00T also achieves speedups of about 1.40× and 4.31× on NVIDIA Thor and A40. With model architectures such as VLA and WAM evolving rapidly and diverging significantly, PhyAI puts ease of use and flexibility first, aiming to reduce the adaptation cost of moving models from research prototypes to efficient inference.
Through a concise API, PhyAI completed adaptation support for MiniCPMRobot on its first day of release and used CUDA Graph for acceleration, raising the inference frame rate from 10.12 Hz to 33.28 Hz. It also applied operator fusion, writing custom fused kernels for MiniCPMRobot in Triton — such as RMSNorm+SiLU gate and conv1d+SiLU+QKV split — to reduce intermediate variable movement and further raise the inference frame rate on NVIDIA H20 to 36.77 Hz.
Bringing on-device AI to every industry
While advancing open-source technology, we are also exploring real-world applications with industry partners.
ModelBest has already partnered with Leju Robotics to launch an agent solution for showroom guidance and campus inspection, and is working with Geely Auto on VLA models and production and warehouse applications, driving continuous validation and iteration of embodied AI technology in real environments.
For ModelBest, exploring physical AGI is part of the same pursuit as the long-term goal of AGI. As robots gradually enter homes, offices, and factories, local perception, inference, and decision-making will become important for guaranteeing real-time performance, stability, and data privacy.
In the future, we will continue to invest in the embodied field, and we will keep driving continuous validation of models in real scenarios through open source and industry collaboration, so that physical intelligence can enter the real world more reliably and safely.