'Address boundary error' when training act policy with aloha_sim_insertion_human

[julyfun]

System Info

- `lerobot` version: 0.1.0
- Platform: Linux-5.15.153.1-microsoft-standard-WSL2+-x86_64-with-glibc2.35
- Python version: 3.10.16
- Huggingface_hub version: 0.27.1
- Dataset version: 3.2.0
- Numpy version: 2.0.2
- PyTorch version (GPU?): 2.6.0+cu124 (True)
- Cuda version: 12040
- Using GPU in script?: <fill in>
(Lerobot on branch master, using cuda 12.4. As in `pyproject.toml` requires torchvision `>=0.21.0`, only pytorch `2.6.0` is compatible)

Information

• One of the scripts in the examples/ folder of LeRobot
• My own task or dataset (give details below)

Reproduction

python lerobot/scripts/train.py \
          --policy.type=act \
          --dataset.repo_id=lerobot/aloha_sim_insertion_human \
          --env.type=aloha \
          --env.task=AlohaInsertion-v0 \
          --log_freq=25 \
          --save_freq=100 

INFO 2025-02-20 20:08:24 ts/train.py:143 step:20K smpl:160K ep:320 epch:6.39 loss:0.073 grdn:6.679 lr:1.0e-05 updt_s:0.081 data_s:0.000
INFO 2025-02-20 20:08:26 ts/train.py:252 Eval policy at step 20000
Stepping through eval batches: 0%| | 0/1 [00:00<?, ?it/s]fish: Job 1, 'python lerobot/scripts/train.py…' terminated by signal --policy.type=act \ ( --dataset.repo_id=lerobot/a…)
fish: Job --env.type=aloha , ' --env.task=AlohaInsertion-v…' terminated by signal --log_freq=25 \ ( --save_freq=100)
fish: Job SIGSEGV, 'Address boundary error' terminated by signal ()

The error appears in the evaluation process after 20K training steps. In bash it shows Segmentation fault (core dumped)

Expected behavior

Evaluation goes to 100%.