Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers

Authors: Lirui Wang, Xinlei Chen, Jialiang Zhao, Kaiming He

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments to investigate the scaling behaviors of training objectives, to the extent of 52 datasets. HPTs outperform several baselines and enhance the fine-tuned policy performance by over 20% on unseen tasks in multiple simulator benchmarks and real-world settings.
Researcher Affiliation Collaboration Lirui Wang1 Xinlei Chen2 Jialiang Zhao1 Kaiming He1 1MIT CSAIL 2Meta, FAIR
Pseudocode No The paper describes the architecture and training process in detail with diagrams and text, but it does not include a dedicated section or figure labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes As an attempt to scale heterogeneous pre-training, our code and weights are open-sourced, and we hope that HPT can shed some light on learning robot representations from heterogeneous embodiments and tasks. ... 2https://github.com/liruiw/HPT and https://github.com/liruiw/lerobot
Open Datasets Yes We use 27 robot teleoperation datasets, including a subset of the recently public Open-X Embodiment dataset [14] as the training corpus. ... In total, we use a subset of 42 datasets in the Open-X Embodiment dataset [14], including the recent Droid [76] dataset.
Dataset Splits Yes We use a maximum of 1000 trajectories from each dataset and a total number of 16k trajectories, and a held-out validation dataset with a maximum 200 trajectories per data source.
Hardware Specification Yes The inference time during transfer on an RTX 3070 GPU is 47Hz for HPT-base and 19Hz for HPT-XL... The compute resources for these pre-training experiments range from 8 V-100s to 128 V-100s... We train with batch size 256 on a single NVIDIA RTX 2080Ti GPU for 20000 iterations.
Software Dependencies No The paper mentions software components like 'Res Net18' and 'T5' with citations, but does not provide specific version numbers for these or other key software dependencies like PyTorch, TensorFlow, or CUDA.
Experiment Setup Yes The training uses a batch size of 256 for 80k iterations... We train HPT with Adam W [47] optimizer with a weight decay ratio 0.05, and a base learning rate of 0.0002 with a cosine learning rate schedule with warmups and dropouts.