Real-World Deployment and Evaluation

[P1star]

Hi! Thanks for releasing this excellent work.

I have a few questions regarding the deployment and real-world evaluation of InternVLA-A1.5:

1. Inference latency / control frequency
   What is the end-to-end inference latency of InternVLA-A1.5 during real-world deployment?
   What control frequency (Hz) does the robot operate at?
2. Hardware setup
   What hardware was used for real-world inference (GPU, CPU, etc.)?
   Were the reported experiments conducted on a single GPU or multiple GPUs?
3. Dynamic task evaluation
   Have you evaluated InternVLA-A1.5 on dynamic manipulation tasks, such as moving objects or changing environments?
   If so, could you share the task settings and the corresponding success rates or any observations?

These details would be very helpful for reproducing the system and understanding its real-world deployment capabilities.

Thank you for your time!

[mahaoxiang822]

Hi,
Thanks for your attention!

1. Inference latency: When using the optimized backend, and deploying on an NVIDIA 5090 GPU, InternVLA-A1.5 uess about 0.1s for once inference, which output an action chunk with size 50.
2. All real-word experiments are deployed on an NVIDIA 5090 GPU. Most of the simulation experiments is conducted on the 4090 gpu cluster.
3. Yes, please see our results on the DOMINO benchmark.

[P1star]

Thank you very much for your reply!

Sorry that I did not make my third question clear enough. I have seen the excellent performance of InternVLA-A1.5 on the DOMINO simulation benchmark.

What I would like to further ask is whether InternVLA-A1.5 has also been evaluated on real-world dynamic manipulation tasks, such as conveyor-belt sorting, moving-object grasping, or other scenarios where the target object or environment changes during execution.

[mahaoxiang822]

Thanks for the clarification!

We haven't systematically evaluated the success rate on these specific dynamic tasks on real robots yet. However, I believe InternVLA-A1.5 has a high probability of accomplishing them, provided that the data collection setup and quality are good.