Missing eval step limits & undocumented pi0.5 baseline training config block reproduction of the Table 2 baseline

[hhh0108]

Hi EventVLA authors,

Thank you very much for releasing EventVLA, RoboTwin-MeM benchmark, checkpoints, and training/evaluation code.

We are currently reproducing EventVLA, and we encounter the following problems:

1. Missing eval step limits

• The released code does not include _eval_step_limit.yml. In envs/_base_task.py any task not listed falls back to step_lim = 1000.
  can you publish the exact per-task eval step limits used to produce the Table 2 results (or ship _eval_step_limit.yml).

we wonder if this mismatch cause Pick Objects in Order cannot reproduce Table 2 result(50% vs 90%)

2. Undocumented baselines training config

The paper reports the pi0.5 baseline at ~20% success on rearrange_blocks_hard, but the training recipe for this baseline is not documented anywhere in the release. The following are unspecified:

• number of training steps
• global batch size
• LoRA vs full fine-tuning
• action horizon and the number of chunk steps executed open-loop per inference at eval

Our reproduction attempt — pi0.5 (PaliGemma-2B + 300M action-expert, LoRA), global batch 32, 30k steps, 50 demos, action horizon 50, executing the full 50-step chunk per inference — gives on rearrange_blocks_hard:

• 0/30 success, average max-reward ≈ 0.28.

Could you publish the pi0.5 baseline training config (steps, batch size, LoRA/full) and the eval protocol (step limit, chunk/replan length, number of trials) used for Table 2.