Adaptation and Generalisation of a Vision-Language-Action Model on a New Embodiment

π₀.₅ fine-tuning, deployment, and generalisation testing on a Universal Robots UR5e (openpi fork).

Vision-language-action (VLA) models promise a single learned policy that adapts to new embodiments from a handful of demonstrations. This work tests that claim: Physical Intelligence's π₀.₅ is fine-tuned on a Universal Robots UR5e for a tabletop pick-and-place task, and the resulting policy is then tested across 17 out-of-distribution conditions using the ⋆-Gen taxonomy to map where it generalises and where it stops working. Adapting the model to the new arm was non-trivial: zero-shot failed and a working policy emerged only after 87 training trials, and generalisation off the training distribution was narrow.

87 runs	80 %	40 %
to a working policy	in-distribution task progress	average across 17 OOD conditions

Results

Adaptation was non-trivial. Zero-shot deployment of the base checkpoint produced no usable behaviour on the UR5e, despite UR5e data being part of the π₀.₅ pre-training mix. Reaching a working policy required 87 training trials across the model variant, fine-tuning method, dataset size, training steps, batch size, learning rate, warm-up, normalisation source, action transform, and inference horizon (full search grid in ur5/docs/training.md). The on-robot working window was narrow in training time: task progress peaked around step 150, and by step 210 the out-of-distribution score had collapsed while the training loss kept falling, so the loss does not reveal the overfitting and each checkpoint has to be checked on the robot. Trial duration also fell on the later checkpoints, as the arm moved more directly to positions it had seen.

_{Across training steps on the 10-episode dataset. Left: in-distribution vs out-of-distribution task progress against training loss; the OOD score collapses at step 210 while the loss keeps falling. Right: trial duration by stage.}

Dataset size and inference horizon mattered too. More data did not help: task progress was best at 10 episodes and dropped with larger datasets. At deployment, the bridge executes K = 6 of the H = 15 actions π₀.₅ predicts per call; too few steps disrupt continuous motions such as the grasp and too many follow stale actions, so both K = 3 and K = 15 collapsed to 25 % task progress (per-experiment notes in ur5/docs/experiments.md).

_{Average task progress against actions executed per inference call (K) at the step-150 checkpoint, with K = 15 the full predicted chunk. Best at roughly one third to one half of H.}

Generalisation was narrow. 85 trials across 17 ⋆-Gen conditions; no trial completed release.

_{Task progress across 17 out-of-distribution conditions. Dashed line is the in-distribution baseline (80 %).}

• Held up. Different gripper (Robotiq 2F-85), raised block, added clutter, and a white block with a matching instruction: all around 75 %, essentially the in-distribution baseline.
• Degraded but partial. New shoulder camera viewpoints (25 to 65 %), different object shapes (45 to 65 %), screwdrivers replacing the block (35 %).
• Broke completely. Either camera covered (0 %), different robot arm (0 %), and a language-conflict scene: "pick up the white block" with the trained blue block also in frame went to the blue block on 4 of 5 trials.

Two specifics worth flagging: the wrist camera was markedly more sensitive than the shoulder camera (a 2 cm wrist shift dropped the policy to the reach stage, while a symmetric back-right shoulder shift still reached transport), and language following was weak whenever a trained object was also visible in the scene.

Discussion

Memorisation, not generalisation. With this little data the policy memorises far more than it generalises. In-distribution task progress of 80 % against about 40 % across the OOD conditions is close to the roughly 42 % an untuned zero-shot policy reached on matched hardware in independent testing, so fine-tuning at this scale improves the trained condition rather than the general ability to solve the task. The policy treats the camera image as a fixed coordinate frame and does not appear to infer scene geometry, the same viewpoint sensitivity seen elsewhere, where accuracy dropped from 94 % to 16 % under viewpoint changes and from 90 % to 0 % under object-position changes. Language following is weak: when the instruction names one object but a trained object is also in frame, the policy follows the visual prior. The π₀.₅ pre-training data is not released, so co-training was not possible and adaptation relied only on the new demonstrations.

Practical takeaways.

• The safe placement zone is small, so do not expect a fine-tuned policy to work outside the position range its data covered.
• A different gripper works, but a different arm does not.
• Match the camera viewpoint to the training data closely, especially for the wrist camera.
• At small data scales, a narrow dataset on the deployment condition is more useful than a varied one.

Setup

UR5e arm, Robotiq Hand-E gripper, two RealSense cameras (one shoulder-mounted at the back-left of the workspace, one on the wrist). π₀.₅ fine-tuned on 10 tele-operated episodes of pick-and-place. Evaluation followed the ⋆-Gen taxonomy: one factor changed at a time, with the rest of the setup held close to training.

Model	π₀.₅
Fine-tune	Full, 150 steps, batch 16
Action transform	Absolute joint positions, converted to deltas at train/inference
Normalisation	Reloaded base UR5e stats
Inference horizon	K = 6 of H = 15 predicted actions per call
Observation	Shoulder camera (back-left) + wrist camera, 2-camera input

_{12 of the 17 tested conditions, grouped by ⋆-Gen category: visual (blue), visual-behavioural (green), visual-semantic (yellow), visual-semantic-behavioural (purple).}

Code

Doc	What's there
ur5/docs/quickstart.md	End-to-end workflow: setup, record, train, deploy
ur5/docs/training.md	Training configs, norm stats, lessons learned
ur5/docs/deployment.md	Docker policy server and robot bridge
ur5/docs/experiments.md	Per-experiment lab notebook

Where every UR5 change lives in the codebase

UR5 code is not all under ur5/. The directory holds the new robot-side code, but the fork also touches src/openpi/ (data loader, transforms, training loop, serving), examples/, scripts/, and the dependency manifests.

1. UR5 hardware and runtime: ur5/ (new directory)

Path	Purpose
ur5/utils/pi0_bridge_ur5_headless.py	Inference bridge: RealSense capture + RTDE robot control + Robotiq gripper + websocket policy client
ur5/utils/rtde_utils.py	Wrappers around UR5e RTDE control / receive interfaces
ur5/utils/robotiq_gripper.py	Robotiq Hand-E gripper control via URCap socket (port 63352)
ur5/defaults.py	Hardware defaults (robot IP, camera serials, joint limits)
ur5/scripts/	Recording, conversion, evaluation, and diagnostics scripts
ur5/test/	Hardware sanity tests (camera list, dual-camera preview)
ur5/docker/serve_policy_robot.Dockerfile	GPU Docker image for combined policy server + robot bridge
ur5/docs/	All thesis-side documentation

2. UR5 example app: examples/ur5/

Path	Status	Purpose
examples/ur5/main.py	new	Example inference entry point
examples/ur5/env.py	new	Example environment wrapper
examples/ur5/real_env.py	new	Real-hardware env (cameras + RTDE)
examples/ur5/README.md	modified	UR5 example overview

3. UR5 policy code: src/openpi/policies/

Path	Status	Purpose
src/openpi/policies/ur5_policy.py	new	UR5Inputs / UR5Outputs transforms (image normalization, joint+gripper packing/slicing)
src/openpi/policies/ur5_policy_test.py	new	Unit tests for the UR5 transforms

4. Training configs: src/openpi/training/config.py (modified)

Adds LeRobotUR5DataConfig plus the UR5 training configs: pi0_ur5, pi05_ur5, pi05_ur5_lora, pi05_ur5_blueblock10, pi0_fast_ur5. Key knobs added: use_delta_action_transform, gripper_oversample_factor, dataset asset-id overrides.

5. Serving and websocket client

Path	What changed
src/openpi/serving/websocket_policy_server.py	keepalive pings disabled (inference can take longer than the default), and policy.infer runs in a thread so the server keeps servicing control frames
packages/openpi-client/src/openpi_client/websocket_client_policy.py	matching client-side adjustments

---

This repository is a research fork of Physical Intelligence's OpenPI. The original project README is collapsed below.

Upstream openpi README (model checkpoints, requirements, installation, fine-tuning recipes, PyTorch support, troubleshooting)

openpi

openpi holds open-source models and packages for robotics, published by the Physical Intelligence team.

Currently, this repo contains three types of models:

• the π₀ model, a flow-based vision-language-action model (VLA).
• the π₀-FAST model, an autoregressive VLA, based on the FAST action tokenizer.
• the π₀.₅ model, an upgraded version of π₀ with better open-world generalization trained with knowledge insulation. Note that, in this repository, we currently only support the flow matching head for both $\pi_{0.5}$ training and inference.

For all models, we provide base model checkpoints, pre-trained on 10k+ hours of robot data, and examples for using them out of the box or fine-tuning them to your own datasets.

This is an experiment: $\pi_0$ was developed for our own robots, which differ from the widely used platforms such as ALOHA and DROID, and though we are optimistic that researchers and practitioners will be able to run creative new experiments adapting $\pi_0$ to their own platforms, we do not expect every such attempt to be successful. All this is to say: $\pi_0$ may or may not work for you, but you are welcome to try it and see!

Updates

• [Sept 2025] We released PyTorch support in openpi.
• [Sept 2025] We released pi05, an upgraded version of pi0 with better open-world generalization.
• [Sept 2025]: We have added an improved idle filter for DROID training.
• [Jun 2025]: We have added instructions for using openpi to train VLAs on the full DROID dataset. This is an approximate open-source implementation of the training pipeline used to train pi0-FAST-DROID.

Requirements

To run the models in this repository, you will need an NVIDIA GPU with at least the following specifications. These estimations assume a single GPU, but you can also use multiple GPUs with model parallelism to reduce per-GPU memory requirements by configuring fsdp_devices in the training config. Please also note that the current training script does not yet support multi-node training.

Mode	Memory Required	Example GPU
Inference	> 8 GB	RTX 4090
Fine-Tuning (LoRA)	> 22.5 GB	RTX 4090
Fine-Tuning (Full)	> 70 GB	A100 (80GB) / H100

The repo has been tested with Ubuntu 22.04, we do not currently support other operating systems.

Installation

When cloning this repo, make sure to update submodules:

git clone --recurse-submodules git@github.com:Physical-Intelligence/openpi.git

# Or if you already cloned the repo:
git submodule update --init --recursive

We use uv to manage Python dependencies. See the uv installation instructions to set it up. Once uv is installed, run the following to set up the environment:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

NOTE: GIT_LFS_SKIP_SMUDGE=1 is needed to pull LeRobot as a dependency.

Docker: As an alternative to uv installation, we provide instructions for installing openpi using Docker. If you encounter issues with your system setup, consider using Docker to simplify installation. See Docker Setup for more details.

Model Checkpoints

Base Models

We provide multiple base VLA model checkpoints. These checkpoints have been pre-trained on 10k+ hours of robot data, and can be used for fine-tuning.

Model	Use Case	Description	Checkpoint Path
$\pi_0$	Fine-Tuning	Base π₀ model for fine-tuning	gs://openpi-assets/checkpoints/pi0_base
$\pi_0$-FAST	Fine-Tuning	Base autoregressive π₀-FAST model for fine-tuning	gs://openpi-assets/checkpoints/pi0_fast_base
$\pi_{0.5}$	Fine-Tuning	Base π₀.₅ model for fine-tuning	gs://openpi-assets/checkpoints/pi05_base

Fine-Tuned Models

We also provide "expert" checkpoints for various robot platforms and tasks. These models are fine-tuned from the base models above and intended to run directly on the target robot. These may or may not work on your particular robot. Since these checkpoints were fine-tuned on relatively small datasets collected with more widely available robots, such as ALOHA and the DROID Franka setup, they might not generalize to your particular setup, though we found some of these, especially the DROID checkpoint, to generalize quite broadly in practice.

Model	Use Case	Description	Checkpoint Path
$\pi_0$-FAST-DROID	Inference	$\pi_0$-FAST model fine-tuned on the DROID dataset: can perform a wide range of simple table-top manipulation tasks 0-shot in new scenes on the DROID robot platform	gs://openpi-assets/checkpoints/pi0_fast_droid
$\pi_0$-DROID	Fine-Tuning	$\pi_0$ model fine-tuned on the DROID dataset: faster inference than $\pi_0$-FAST-DROID, but may not follow language commands as well	gs://openpi-assets/checkpoints/pi0_droid
$\pi_0$-ALOHA-towel	Inference	$\pi_0$ model fine-tuned on internal ALOHA data: can fold diverse towels 0-shot on ALOHA robot platforms	gs://openpi-assets/checkpoints/pi0_aloha_towel
$\pi_0$-ALOHA-tupperware	Inference	$\pi_0$ model fine-tuned on internal ALOHA data: can unpack food from a tupperware container	gs://openpi-assets/checkpoints/pi0_aloha_tupperware
$\pi_0$-ALOHA-pen-uncap	Inference	$\pi_0$ model fine-tuned on public ALOHA data: can uncap a pen	gs://openpi-assets/checkpoints/pi0_aloha_pen_uncap
$\pi_{0.5}$-LIBERO	Inference	$\pi_{0.5}$ model fine-tuned for the LIBERO benchmark: gets state-of-the-art performance (see LIBERO README)	gs://openpi-assets/checkpoints/pi05_libero
$\pi_{0.5}$-DROID	Inference / Fine-Tuning	$\pi_{0.5}$ model fine-tuned on the DROID dataset with knowledge insulation: fast inference and good language-following	gs://openpi-assets/checkpoints/pi05_droid

By default, checkpoints are automatically downloaded from gs://openpi-assets and are cached in ~/.cache/openpi when needed. You can overwrite the download path by setting the OPENPI_DATA_HOME environment variable.

Running Inference for a Pre-Trained Model

Our pre-trained model checkpoints can be run with a few lines of code (here our $\pi_0$-FAST-DROID model):

from openpi.training import config as _config
from openpi.policies import policy_config
from openpi.shared import download

config = _config.get_config("pi05_droid")
checkpoint_dir = download.maybe_download("gs://openpi-assets/checkpoints/pi05_droid")

# Create a trained policy.
policy = policy_config.create_trained_policy(config, checkpoint_dir)

# Run inference on a dummy example.
example = {
    "observation/exterior_image_1_left": ...,
    "observation/wrist_image_left": ...,
    ...
    "prompt": "pick up the fork"
}
action_chunk = policy.infer(example)["actions"]

You can also test this out in the example notebook.

We provide detailed step-by-step examples for running inference of our pre-trained checkpoints on DROID and ALOHA robots.

Remote Inference: We provide examples and code for running inference of our models remotely: the model can run on a different server and stream actions to the robot via a websocket connection. This makes it easy to use more powerful GPUs off-robot and keep robot and policy environments separate.

Test inference without a robot: We provide a script for testing inference without a robot. This script will generate a random observation and run inference with the model. See here for more details.

Fine-Tuning Base Models on Your Own Data

We will fine-tune the $\pi_{0.5}$ model on the LIBERO dataset as a running example for how to fine-tune a base model on your own data. We will explain three steps:

1. Convert your data to a LeRobot dataset (which we use for training)
2. Defining training configs and running training
3. Spinning up a policy server and running inference

1. Convert your data to a LeRobot dataset

We provide a minimal example script for converting LIBERO data to a LeRobot dataset in examples/libero/convert_libero_data_to_lerobot.py. You can easily modify it to convert your own data! You can download the raw LIBERO dataset from here, and run the script with:

uv run examples/libero/convert_libero_data_to_lerobot.py --data_dir /path/to/your/libero/data

Note: If you just want to fine-tune on LIBERO, you can skip this step, because our LIBERO fine-tuning configs point to a pre-converted LIBERO dataset. This step is merely an example that you can adapt to your own data.

2. Defining training configs and running training

To fine-tune a base model on your own data, you need to define configs for data processing and training. We provide example configs with detailed comments for LIBERO below, which you can modify for your own dataset:

• LiberoInputs and LiberoOutputs: Defines the data mapping from the LIBERO environment to the model and vice versa. Will be used for both, training and inference.
• LeRobotLiberoDataConfig: Defines how to process raw LIBERO data from LeRobot dataset for training.
• TrainConfig: Defines fine-tuning hyperparameters, data config, and weight loader.

We provide example fine-tuning configs for π₀, π₀-FAST, and π₀.₅ on LIBERO data.

Before we can run training, we need to compute the normalization statistics for the training data. Run the script below with the name of your training config:

uv run scripts/compute_norm_stats.py --config-name pi05_libero

Now we can kick off training with the following command (the --overwrite flag is used to overwrite existing checkpoints if you rerun fine-tuning with the same config):

XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run scripts/train.py pi05_libero --exp-name=my_experiment --overwrite

The command will log training progress to the console and save checkpoints to the checkpoints directory. You can also monitor training progress on the Weights & Biases dashboard. For maximally using the GPU memory, set XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 before running training -- this enables JAX to use up to 90% of the GPU memory (vs. the default of 75%).

Note: We provide functionality for reloading normalization statistics for state / action normalization from pre-training. This can be beneficial if you are fine-tuning to a new task on a robot that was part of our pre-training mixture. For more details on how to reload normalization statistics, see the norm_stats.md file.

3. Spinning up a policy server and running inference

Once training is complete, we can run inference by spinning up a policy server and then querying it from a LIBERO evaluation script. Launching a model server is easy (we use the checkpoint for iteration 20,000 for this example, modify as needed):

uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_libero --policy.dir=checkpoints/pi05_libero/my_experiment/20000

This will spin up a server that listens on port 8000 and waits for observations to be sent to it. We can then run an evaluation script (or robot runtime) that queries the server.

For running the LIBERO eval in particular, we provide (and recommend using) a Dockerized workflow that handles both the policy server and the evaluation script together. See the LIBERO README for more details.

If you want to embed a policy server call in your own robot runtime, we have a minimal example of how to do so in the remote inference docs.

More Examples

We provide more examples for how to fine-tune and run inference with our models on the ALOHA platform in the following READMEs:

• ALOHA Simulator
• ALOHA Real
• UR5

PyTorch Support

openpi now provides PyTorch implementations of π₀ and π₀.₅ models alongside the original JAX versions! The PyTorch implementation has been validated on the LIBERO benchmark (both inference and finetuning). A few features are currently not supported (this may change in the future):

• The π₀-FAST model
• Mixed precision training
• FSDP (fully-sharded data parallelism) training
• LoRA (low-rank adaptation) training
• EMA (exponential moving average) weights during training

Setup

1. Make sure that you have the latest version of all dependencies installed: uv sync

2. Double check that you have transformers 4.53.2 installed: uv pip show transformers

3. Apply the transformers library patches:

   cp -r ./src/openpi/models_pytorch/transformers_replace/* .venv/lib/python3.11/site-packages/transformers/

This overwrites several files in the transformers library with necessary model changes: 1) supporting AdaRMS, 2) correctly controlling the precision of activations, and 3) allowing the KV cache to be used without being updated.

WARNING: With the default uv link mode (hardlink), this will permanently affect the transformers library in your uv cache, meaning the changes will survive reinstallations of transformers and could even propagate to other projects that use transformers. To fully undo this operation, you must run uv cache clean transformers.

Converting JAX Models to PyTorch

To convert a JAX model checkpoint to PyTorch format:

uv run examples/convert_jax_model_to_pytorch.py \
    --checkpoint_dir /path/to/jax/checkpoint \
    --config_name <config name> \
    --output_path /path/to/converted/pytorch/checkpoint

Running Inference with PyTorch

The PyTorch implementation uses the same API as the JAX version - you only need to change the checkpoint path to point to the converted PyTorch model:

from openpi.training import config as _config
from openpi.policies import policy_config
from openpi.shared import download

config = _config.get_config("pi05_droid")
checkpoint_dir = "/path/to/converted/pytorch/checkpoint"

# Create a trained policy (automatically detects PyTorch format)
policy = policy_config.create_trained_policy(config, checkpoint_dir)

# Run inference (same API as JAX)
action_chunk = policy.infer(example)["actions"]

Policy Server with PyTorch

The policy server works identically with PyTorch models - just point to the converted checkpoint directory:

uv run scripts/serve_policy.py policy:checkpoint \
    --policy.config=pi05_droid \
    --policy.dir=/path/to/converted/pytorch/checkpoint

Finetuning with PyTorch

To finetune a model in PyTorch:

1. Convert the JAX base model to PyTorch format:

   uv run examples/convert_jax_model_to_pytorch.py \
       --config_name <config name> \
       --checkpoint_dir /path/to/jax/base/model \
       --output_path /path/to/pytorch/base/model

2. Specify the converted PyTorch model path in your config using pytorch_weight_path

3. Launch training using one of these modes:

# Single GPU training:
uv run scripts/train_pytorch.py <config_name> --exp_name <run_name> --save_interval <interval>

# Example:
uv run scripts/train_pytorch.py debug --exp_name pytorch_test
uv run scripts/train_pytorch.py debug --exp_name pytorch_test --resume  # Resume from latest checkpoint

# Multi-GPU training (single node):
uv run torchrun --standalone --nnodes=1 --nproc_per_node=<num_gpus> scripts/train_pytorch.py <config_name> --exp_name <run_name>

# Example:
uv run torchrun --standalone --nnodes=1 --nproc_per_node=2 scripts/train_pytorch.py pi0_aloha_sim --exp_name pytorch_ddp_test
uv run torchrun --standalone --nnodes=1 --nproc_per_node=2 scripts/train_pytorch.py pi0_aloha_sim --exp_name pytorch_ddp_test --resume

# Multi-Node Training:
uv run torchrun \
    --nnodes=<num_nodes> \
    --nproc_per_node=<gpus_per_node> \
    --node_rank=<rank_of_node> \
    --master_addr=<master_ip> \
    --master_port=<port> \
    scripts/train_pytorch.py <config_name> --exp_name=<run_name> --save_interval <interval>

Precision Settings

JAX and PyTorch implementations handle precision as follows:

JAX:

1. Inference: most weights and computations in bfloat16, with a few computations in float32 for stability
2. Training: defaults to mixed precision: weights and gradients in float32, (most) activations and computations in bfloat16. You can change to full float32 training by setting dtype to float32 in the config.

PyTorch:

1. Inference: matches JAX -- most weights and computations in bfloat16, with a few weights converted to float32 for stability
2. Training: supports either full bfloat16 (default) or full float32. You can change it by setting pytorch_training_precision in the config. bfloat16 uses less memory but exhibits higher losses compared to float32. Mixed precision is not yet supported.

With torch.compile, inference speed is comparable between JAX and PyTorch.

Troubleshooting

We will collect common issues and their solutions here. If you encounter an issue, please check here first. If you can't find a solution, please file an issue on the repo (see here for guidelines).

Issue	Resolution
uv sync fails with dependency conflicts	Try removing the virtual environment directory (rm -rf .venv) and running uv sync again. If issues persist, check that you have the latest version of uv installed (uv self update).
Training runs out of GPU memory	Make sure you set XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 (or higher) before running training to allow JAX to use more GPU memory. You can also use --fsdp-devices <n> where <n> is your number of GPUs, to enable fully-sharded data parallelism, which reduces memory usage in exchange for slower training (the amount of slowdown depends on your particular setup). If you are still running out of memory, you may want to consider disabling EMA.
Policy server connection errors	Check that the server is running and listening on the expected port. Verify network connectivity and firewall settings between client and server.
Missing norm stats error when training	Run scripts/compute_norm_stats.py with your config name before starting training.
Dataset download fails	Check your internet connection. For HuggingFace datasets, ensure you're logged in (huggingface-cli login).
CUDA/GPU errors	Verify NVIDIA drivers are installed correctly. For Docker, ensure nvidia-container-toolkit is installed. Check GPU compatibility. You do NOT need CUDA libraries installed at a system level --- they will be installed via uv. You may even want to try uninstalling system CUDA libraries if you run into CUDA issues, since system libraries can sometimes cause conflicts.
Import errors when running examples	Make sure you've installed all dependencies with uv sync. Some examples may have additional requirements listed in their READMEs.
Action dimensions mismatch	Verify your data processing transforms match the expected input/output dimensions of your robot. Check the action space definitions in your policy classes.
Diverging training loss	Check the q01, q99, and std values in norm_stats.json for your dataset. Certain dimensions that are rarely used can end up with very small q01, q99, or std values, leading to huge states and actions after normalization. You can manually adjust the norm stats as a workaround.