[Community] MiniCPM-V 4.6 integrated into StarVLA for robot manipulation

[shaohua-pan]

Hi MiniCPM-V team,

I recently integrated MiniCPM-V 4.6 as a new vision-language backbone for StarVLA, a Vision-Language-Action framework for robot manipulation.

PR: starVLA/starVLA#354

The integration adds:

• MiniCPM-V 4.6 VLM wrapper for StarVLA
• PI / GR00T-style VLA framework support
• LIBERO training and evaluation scripts
• Example configs for 8-GPU training

Using MiniCPM-V 4.6, we trained and evaluated on the LIBERO benchmark and got the following results:

Benchmark	Success Rate
LIBERO-Spatial	94.0%
LIBERO-Object	98.0%
LIBERO-Goal	98.0%
LIBERO-10	92.4%
Overall	95.6%

Training setup:

• Backbone: openbmb/MiniCPM-V-4.6
• Effective batch size: 128
• Max training steps: 80k
• Attention implementation: flash_attention_2
• No modules are frozen by default (FREEZE_MODULES=""), so MiniCPM-V is trainable unless overridden.

MiniCPM-V 4.6 looks like a strong lightweight VLM backbone for embodied AI / VLA-style robot learning, especially when compared with larger VLM backbones.

As a next step, we are also considering testing this MiniCPM-V-based StarVLA model on a real SO-101 robot setup, to further evaluate its sim-to-real potential beyond LIBERO simulation benchmarks.

Thanks for releasing MiniCPM-V! Feedback or suggestions from the MiniCPM-V team would be very welcome.

[tc-mb]

Hi @shaohua-pan, this is really exciting — thanks for taking the time to share the integration and benchmark results in such detail.

The LIBERO numbers are impressive, especially for a model of this size. Seeing MiniCPM-V 4.6 perform as a lightweight VLM backbone for robot manipulation is exactly the kind of downstream use case we'd hoped to see. The sim-to-real test on the SO-101 would be a great next step — definitely keep us posted on how that goes.

Happy to help if you run into any model-side questions along the way. Feel free to open an issue or reach out here anytime.