OpenBMB Open-Sources MiniCPM-Robot: 1.5B VLA Beats π0.5 by 5x on Memory Benchmarks
Source: OpenBMB / Xinzhiyuan / Chinaz URL: https://x.com/OpenBMB/status/2078839529591759025 Date: July 19, 2026 (WAIC on-site)
What Happened
At WAIC 2026, Tsinghua-affiliated unicorn OpenBMB open-sourced its first embodied AI model family, MiniCPM-Robot, releasing three components at once:
- MiniCPM-RobotManip: a 1.5B-parameter general vision-language-action (VLA) model for robot manipulation.
- MiniCPM-RobotTrack: a 0.9B-parameter end-to-end visual instruction-tracking model for target following.
- PhyAI: a high-performance inference framework designed for embodied models.
- Real-robot generalization at 1.5B — benchmark numbers look great, but RMBench is a new benchmark; performance under complex lighting, occlusion, and deformable objects remains to be seen.
- Open source ≠ ecosystem — Llama took two years to build an ecosystem; VLA models involve hardware adaptation, data collection, and sim-to-real transfer, making ecosystem building far harder than for LLMs.
- Training data sources not fully disclosed — a 1.5B model typically needs millions of trajectories; their provenance and compliance aren't stated in the GitHub repo.
- OpenBMB has no robot mass-production capability — open-sourcing the model is step one; running it on 10,000 robots requires downstream integrators, and whether OpenBMB does this itself or via ecosystem partners is unclear.
- Tweet: https://x.com/OpenBMB/status/2078839529591759025
- Xinzhiyuan: https://www.163.com/dy/article/L27AB1VS0511ABV6.html
- Chinaz: https://www.chinaz.com/ainews/29702.shtml
- GitHub: https://github.com/OpenBMB/MiniCPM-Robot
- Hugging Face: https://huggingface.co/openbmb/MiniCPM-RobotManip
All weights and code are available on GitHub (OpenBMB/MiniCPM-Robot), with model weights also on Hugging Face — no review, no API quotas, no NDA, a rare level of openness in the 2026 embodied AI space.
Deep Dive
The Key Isn't 1.5B — It's Memory
The pain point of embodied intelligence isn't perception; it's contextual memory.
Most mainstream VLAs (π0.5, RT-2, OpenVLA) are essentially "single-frame reactive": they see the current frame, decide the next action, then discard everything prior. Ask a robot to press a button five times, and by the second press the model has already forgotten the first.
Why not add history? Compute. Each extra historical frame causes near-exponential compute growth; forcing memory in makes the robot sluggish and unusable.
MiniCPM-RobotManip's solution is quintessentially ModelBest: port the visual token compression techniques of MiniCPM-V 4.6 into robotics. The results:
| Metric | π0.5 (mainstream) | MiniCPM-RobotManip | Multiple | |---|---|---|---| | RMBench context-memory score | 10.4 | 53.5 | 5.1× | | Compute for 60-frame history decisions | 5.0 TFLOPs (single frame) | 3.3 TFLOPs (streaming) | -34% | | Per-decision latency on H100 | 234ms | 120ms | -49% |
The trick is "streaming inference": instead of recomputing historical frames, it incrementally updates a visual token cache, so "the cost of keeping 1 minute of context memory" matches "traditional single-frame reactive" inference. Better performance and lower deployment cost holding simultaneously is rare in the 2026 VLA landscape.
The Tracking Model: 0.9B on a Single Card, Works Even in Network-Free Elevators
MiniCPM-RobotTrack sets open-source SOTA on three tasks:
| Task | MiniCPM-RobotTrack (0.9B) | vs OmTrackVLA | vs TrackVLA++ (closed) | |---|---|---|---| | Single-target tracking (STT) | 89.8 | +7.0pp | +8.8pp | | Dynamic multi-target (DT) | 73.4 | +14.6pp | — | | Ambiguous target (AT) | 80.4 (58.0% success) | +6.5pp | +17.0pp |
It natively supports the Unitree Go2 Edu, running at 5+ Hz with 180ms end-to-end response using the stock camera and onboard compute. It has been demonstrated smoothly tracking in "no-network/weak-network" scenarios like elevators and parking garages — arguably more meaningful than the benchmark numbers: real robot deployment compressed to consumer hardware.
Industry Position: 1.5B Isn't "Small" — It's "Actually Runnable on Real Robots"
| Model | Params | Real-robot requirement | |---|---|---| | π0.5 (Physical Intelligence) | ~3B | H100/A100 | | OpenVLA-7B | 7.8B | Multi-GPU | | RT-2-X | 55B | Massive cluster | | MiniCPM-RobotManip | 1.5B | Consumer GPU |
The Tsinghua lineage has pushed "embodied AI democratization" from paper demos to something individual developers can clone and run on real robots. Once the barrier is kicked open via open source, iteration speed in embodied AI may exceed most expectations.
Why It Matters
1. Memory is the bottleneck of embodiment — everyone else is trying to route around it; ModelBest attacked it head-on with visual token compression. This is genuine hard tech, not a wrapper. 2. Full weights released at 1.5B — rarely seen openness in embodied AI; researchers, robot hobbyists, and startups outside big labs can get started immediately. 3. Real-robot deployment pushed to consumer level — a ~10k-RMB Go2 Edu robot dog plus one consumer GPU can run end-to-end VLA. Embodied AI's "iPhone moment" may be closer than expected. 4. The embodied track is collectively launching in 2026 H2 — Riemann Dynamics' Riemann-1.0 (232k hours of embodied interaction data, dual SOTA in real-robot and simulation) also debuted at the same WAIC.
Risks and Things to Watch
---
One-Line Summary
A 1.5B model beating π0.5 by 5x, 120ms on a single card, 180ms tracking a person on a Unitree Go2 — OpenBMB has dragged embodied AI from "lab demo" to the threshold of "developers clone and run." More worth tracking than LingBot-VA's complex VAE architecture, because it bets directly on real-robot runnability.