Questions about RoboCasa365 support, benchmark hardware/time, and LLM cost

[coolmaoguo]

Hi RPent team,

Thanks for the great work!

I have three questions about RoboCasa365 and the reported benchmark:

1. The README mentions scripts/setup_robocasa.sh, scripts/run_robocasa.sh, and docs/SETUP_ROBOCASA.zh.md, but I cannot find them in the current main branch. Is the RoboCasa365 integration already open-sourced? If not, when is it expected to be released?

2. Could you share the hardware setup and evaluation time for the reported benchmark results? For example, GPU type/count, parallelization setting, time per task/seed, or total evaluation time.

3. Could you provide an approximate LLM API cost for running the benchmark?
   Thanks!

[jx-qiu]

Hi coolmaogo,

1. Robocasa support is planned and will be available in the next few weeks.
2. You will need ~10-12 GiB of free GPU memory to run RPent on LIBERO Pro. Parallelization settings don't affect the results, so choose it according to available GPU memory and MaaS API concurrency. A single LIBERO run typically takes 3-10 minutes to complete, depending on the LLM used.
3. Using Claude Sonnet, each run costs about ~$0.5-2.

[liberow]

Hi RPent team,
I have one more question after reading the Harness VLA paper:

The paper mentions using RLDX-1 (the frozen VLA checkpoint from arXiv:2605.03269) as the vla_act backend for RoboCasa365. I noticed in the README Feature Matrix that RLDX-1 is listed without a ✅ checkmark. Is the RLDX-1 model checkpoint open-sourced? If so, where can we find the weights? If not, are there plans to release it, or will you support other open VLA backends (e.g., π₀.₅) for RoboCasa365 evaluation?

Thanks!

[jx-qiu]

@qurakchin do you know when RLDX-1 will be released?