I trained the model with this data, and the L1 loss went down, but for some reason the model keeps failing the tasks.
If you use this dataset or any part of the conversion pipeline, please cite the original RoboCerebra paper:
@article{han2025robocerebra,
title={RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation},
author={Han, Songhao and Qiu, Boxiang and Liao, Yue and Huang, Siyuan and Gao, Chen and Yan… See the full description on the dataset page:
https://huggingface.co/datasets/Yun5/RoboCerebra_TF.
