Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Research & technology· 日本語で読む

RoboParty releases repository for training, evaluation and policy export for RPO and RP1

RoboParty’s public repository, “roboparty_train,” supports training, evaluation and policy export for RPO and RP1. It is based on Isaac Lab and RSL-RL, with a choice of MJWarp or Isaac Sim PhysX. It includes AMP, BeyondMimic and Parkour workflows, along with robot-specific motion data.

Key facts

  • roboparty_train provides training, evaluation and policy export for RPO and RP1.
  • It lists Isaac Lab 3.0.0 GA and a fork of RSL-RL 5.5.1 as its foundations.
  • Its training methods include AMP, BeyondMimic and Parkour workflows.
  • Motion-conversion examples specify output rates of 120 Hz for AMP and 50 Hz for BeyondMimic.
  • The RP1-Flat training command example specifies “--num_envs=4096.”

What happened

RoboParty has released the GitHub repository “roboparty_train” for training, evaluating and exporting policies for RPO and RP1. It is based on Isaac Lab 3.0.0 GA and a fork of RSL-RL 5.5.1, and provides configurations using either MJWarp or Isaac Sim PhysX. Its training methods include AMP, BeyondMimic and Parkour workflows.

Why it matters

Users of RPO and RP1 can follow procedures in a single repository for training, evaluation and policy export. However, AMP and BeyondMimic use different motion formats and sample rates, while MJWarp and PhysX run in separate environments, so users must convert data as needed and select an execution environment for each method.

Analysis

The repository’s distinguishing feature is that it brings simulation-based training, evaluation and policy export together in the same set of development assets, rather than limiting its scope to robot design. It supports RPO and RP1 environments and offers AMP, BeyondMimic and Parkour, while allowing users to choose between MJWarp and Isaac Sim PhysX for physics simulation. However, the value 4096 is the number of environments specified in the RP1-Flat command example; it is not a measurement of training speed or real-world performance. It is important not to interpret it as a published processing-capacity figure. Whereas the development assets covered by this newspaper in “RP1 mechanical design through training made public on GitHub” made the design and training resources public, this repository provides details on the software for training, evaluation and export. Its role also differs from the deployment side, such as robot connections and inference, described in ROS 2 deployment framework for RPO and RP1. The publicly available components span multiple stages, from design and training to deployment in operating environments; this repository alone does not guarantee real-world operation or consistent compatibility. There are three structural considerations. First, because MJWarp and PhysX use separate Python environments, users must choose a physics backend and match dependencies and execution requirements. Second, AMP and BeyondMimic use different example data formats and output rates, so preparing and converting motion data is a task specific to each training path. Third, policy export brings the steps for passing training results to subsequent stages into scope, but the description alone does not establish compatibility or performance on physical robots. The next questions to examine are which robot configurations and real-world conditions are used to evaluate policies trained with each method, the conditions under which training data are created and the scope of their permitted use, and the execution requirements for exported models. The repository describes procedures and data formats, but does not specify training time, real-world success rates, safety evaluations or the specific hardware configurations supported. The scope of the publicly available development platform should therefore be distinguished from capabilities validated on physical robots.

Details

The repository consists of “robolab,” which includes RoboParty’s Isaac Lab extensions, environments, scripts and robot assets, and a fork of RSL-RL that supports AMP and Parkour policies. MJWarp and PhysX use separate Python environments. The RP1 training example specifies “--num_envs=4096” for the “RP1-Flat” task. This is a command setting, not a measured processing-capacity figure. The included motion data provides PKL files for AMP and NPZ files for BeyondMimic for both RPO and RP1. In conversion examples from HHTools, AMP uses 120 Hz and BeyondMimic 50 Hz; the target robot is specified as “rp1” or “rpo” during conversion. Parkour export requires the corresponding “.onnx.data” files in addition to “0-depth_encoder.onnx” and “actor.onnx.”

Sources

Translated from the Japanese original: /news/roboparty-rpo-to-rp1-muke-no-gakushuu-hyouka-porishii-kakida-287963. The Japanese version is authoritative if they differ.