OpenPI

Backend choice changes the available workflow

The pinned README supports JAX and selected PyTorch models. Its PyTorch section excludes pi0-FAST, LoRA, FSDP, mixed-precision training and EMA weights during training. Therefore, a memory estimate or fine-tuning recipe elsewhere in the overview is not automatically available through both implementations. The PyTorch setup also patches transformers 4.53.2; upstream warns that the default uv hardlink mode can propagate those modifications through the shared cache. Review that setup in an isolated environment before reusing it across projects. [S001]

The requirements section says the current training script lacks multi-node support, while the later PyTorch section supplies a multi-node launch example. That discrepancy needs a backend-specific reading and validation; this profile does not assert one universal multi-node capability. Upstream's stated supported system is Ubuntu 22.04 with an NVIDIA GPU. Its single-GPU estimates are greater than 8 GB for inference, 22.5 GB for LoRA and 70 GB for full fine-tuning, not measured requirements from this review. [S001]

A checkpoint needs its configuration and statistics

The README lists base and task-specific checkpoint paths in the public openpi-assets bucket. They are downloaded and cached by the example utilities when used. Selecting a path is only part of setup: the policy configuration supplies the model and data transformations, and normalization statistics accompany the checkpoint. This review inspected these instructions without downloading model payloads. [S001] [S005]

The normalization guide distinguishes recomputing statistics for new data from reusing statistics for a matching pretraining robot. Reuse requires the same action-space definitions, not merely the same robot brand. It recommends evaluating both choices when adapting a task. The guide's base-model conventions include joint angles in radians and gripper positions from zero for open to one for closed, but DROID uses joint-velocity actions and a 15 Hz control frequency. Those differences must be resolved before sending a predicted action to a robot. [S005]

The listed normalization asset identifiers are specific to the documented base-model routes. Do not infer that every statistic applies unchanged to every newer checkpoint. The README's data-conversion workflow also computes statistics before fine-tuning and describes input/output transformations for the selected environment. [S001] [S005]

Remote inference separates environments, not robot semantics

The remote-inference example runs the policy on a GPU server and uses a lightweight WebSocket client in the robot environment. Its client resizes images, converts them to unsigned 8-bit format and sends policy-specific observation keys, state and a task prompt. It returns an action chunk with horizon and action dimensions. The robot-side integration still has to interpret and execute those actions according to the selected policy's conventions. [S006]

Use the repository's no-robot inference example and the relevant robot or simulator instructions as staged integration checks. That is an editorial starting sequence; this publication has not run inference, training, networking or hardware control and does not report latency or task-success measurements. [S001] [S006]

Separate source, checkpoint and training-data rights

The source root license is Apache-2.0. The repository also includes Gemma Terms of Use, which contain use restrictions and distribution obligations. The inspected materials do not establish a complete license mapping for every published checkpoint, so this profile does not label all weights Apache-2.0 or uniformly open source. [S003] [S008]

Likewise, public fine-tuning examples and dataset links do not establish release of the entire pretraining mixture. The overview describes more than 10,000 hours of training data and references both public and internal task data. Availability and rights for the complete mixture remain unresolved here. Source availability, downloadable checkpoints and full training reproducibility are distinct findings. [S001]