robomimic

Installation and experiment reproduction are separate checks

The pinned installation guide names Linux and macOS, recommends Python 3.8.0 and shows PyTorch 2.0.0 with torchvision 0.15.1. Package metadata has a much broader Python declaration and also pins several learning dependencies, including transformers 4.41.2 and diffusers 0.11.1. These are source-level instructions, not a resolved environment tested here. Use an isolated environment and record the versions that actually install before treating it as reproducible. [S005] [S006]

The guide's small installation check is examples/train_bc_rnn.py --debug. Dataset use can require additional simulators. For the released robosuite demonstration route, the documentation points to robosuite v1.5.1. That is a more specific choice than its general statement of robosuite v1.2-or-newer compatibility. [S005]

The dataset guide explains why the distinction matters: the original CoRL 2021 data used the older offline_study branch and mujoco-py. Newer datasets use robosuite v1.5.1 with the official MuJoCo bindings, and upstream explicitly says learning results may not match the original results exactly. Older reproduction routes are documented through robomimic v0.2.0 or v0.3 for their respective dataset generations. Do not silently substitute one generation when reporting a historical comparison. [S009]

Inspect the data before training

A compatible HDF5 file contains trajectory groups and environment metadata, with actions, observations, rewards and termination signals. Images are stored as unsigned 8-bit arrays in channel-last order. Actions should be normalized between minus one and one; the documented dataset inspection script checks their range. An HDF5 extension alone therefore does not establish compatibility. [S007]

The robosuite conversion route changes a raw demonstration file in place. After that conversion, the data still lacks observations, rewards and termination signals needed for training. The separate state-to-observation step reconstructs selected observations and rewards from the recorded MuJoCo states. Preserve a copy of raw data before an in-place conversion; this is an editorial precaution for keeping the original evidence. [S008]

Generate train-validation splits before extracting observations. Upstream recommends that ordering so every derived HDF5 file inherits the same trajectory filters. Otherwise, comparisons between image and low-dimensional observations can inadvertently use different splits. The framework exposes filter keys for selecting training and validation trajectories explicitly. [S007] [S008]

Observation extraction also chooses camera names, resolution, depth, reward density and termination semantics. Its examples distinguish marking success from marking only the trajectory end. Compression trades smaller storage for decoding work; omitting next observations is appropriate only for algorithms that do not require them. Preserve these choices with the experiment configuration rather than treating every derived dataset as equivalent. [S008]

Keep source and dataset rights separate

The framework's root license is MIT. The separately inspected maintainer-hosted Hugging Face dataset card also labels its collection MIT, but that collection is only part of the broader supported-data ecosystem. The dataset guide says its real-robot releases are not hosted there. Neither license statement establishes the terms of every external dataset supported by the framework. [S003] [S009] [S010]

This review inspected source and documentation and preserved the dataset card revision. It did not download the HDF5 payloads, resolve an installation, replay states, train a policy or measure a reproduction result. [S002] [S010]