Fine-tuning the π0 and π0.5 Models

π0 and π0.5 are Vision-Language-Action (VLA) policies from Physical Intelligence, and their training pipeline comes from OpenPI. They are the openpi-0 and openpi-0.5 models of the OpenPI repository, written officially as π₀ and π₀.₅, registered under the model identifiers pi0 and pi05 with base checkpoints pi0_base and pi05_base. The platform registers the two models under those identifiers. Both support the jax and lerobot training frameworks and take LeRobot v2 and v3 datasets.

Roles and Prerequisites

Roles

Role	Can create a training task	Description
Administrator	Yes	Not restricted by Module Permissions
Project Manager	Requires authorization	An administrator grants Create Training Job in Module Permissions
Annotator	Requires authorization	Same as above
Auditor	Requires authorization	Same as above
Collector	Requires authorization	Same as above

Viewing training tasks and metrics requires only View Training.

Prerequisites

Item	Requirement
Dataset version	LeRobot v2 or v3. The training service reads codebase_version in the dataset's meta/info.json to identify the version and choose the training image
Framework	jax (OpenPI pipeline) or lerobot
GPU memory	About 22.5 GB or more with LoRA enabled on a single GPU; about 70 GB or more for full fine-tuning
Dependency	The platform's model training service is enabled with at least one training location configured; the jax framework uses the OpenPI training image, and the lerobot framework selects a LeRobot training image based on the dataset version
Dataset structure	The directory contains meta/info.json, data/, and videos/
Permission	Creating requires Create Training Job; viewing requires View Training

Training Task Creation and Parameters

Procedure

1. Go to Model → Training and select Create Training Job.
2. Choose a training location: a local GPU or a cloud training location enabled by an administrator.
3. Under Selected Model, select pi0 or pi05.
4. Under Training framework, select jax or lerobot. The framework determines which parameters appear in Training Parameters.
5. Under Training Dataset, choose the data source: Export Records, an uploaded LeRobot dataset, Download Link, or Hugging Face.
6. Set parameters in Training Parameters; the table below lists their meaning and default values.
7. Enter Notes (optional).
8. Select Create Training Job.

Training Parameters

The jax and lerobot frameworks use two separate parameter sets. The jax framework targets the OpenPI training entry point:

Parameter	Description	Default	Valid range
batch_size	Global batch size	1	Integer, ≥ 1
steps	Training steps	10000	Integer, ≥ 1
save_interval	Checkpoint save interval	5000	Integer, ≥ 1 and not greater than steps
learning_rate	Peak learning rate of the cosine decay schedule	2.5e-5	Float, > 0
fsdp_devices	Number of FSDP devices	auto	auto or a positive integer; not greater than the visible GPU count and a divisor of it
ema_decay	EMA decay factor	Not enabled	0 < x ≤ 1; ignored when LoRA is enabled
action_horizon	Length of the action sequence predicted per step	50	Integer, ≥ 1
prompt	Default instruction when the dataset has no task text	Empty	String

The lerobot framework first applies general training parameters:

Parameter	Description	Default	Valid range
batch_size	Batch size	1	Integer, ≥ 1
steps	Training steps	10000	Integer, ≥ 1
seed	Random seed	1000	Integer
num_workers	DataLoader processes	4	Integer, ≥ 1
eval_freq	Evaluation interval	1000	Integer, ≥ 1 and not greater than steps
log_freq	Logging interval	100	Integer, ≥ 1 and not greater than steps
save_freq	Checkpoint save interval	5000	Integer, ≥ 1 and not greater than steps
save_checkpoint	Whether to save checkpoints	true	true / false

It then appends model parameters with the policy. prefix:

Parameter	Description	Default	Valid range
policy.n_obs_steps	Input observation steps	1	Integer
policy.chunk_size	Action prediction length	50	Integer
policy.n_action_steps	Action steps executed per call	50	Integer
policy.max_state_dim	State vector padding dimension	32	Integer
policy.max_action_dim	Action vector padding dimension	32	Integer
policy.num_inference_steps	Denoising steps at inference	10	Integer
policy.dtype	Model data type	bfloat16	bfloat16, float32
policy.tokenizer_max_length	Maximum tokenizer length	48 for pi0, 200 for pi05	Integer
policy.optimizer_lr	Optimizer learning rate	2.5e-5	Float
policy.optimizer_weight_decay	Weight decay	0.0	Float
policy.optimizer_grad_clip_norm	Gradient clipping threshold	1.0	Float
policy.scheduler_warmup_steps	Learning rate warmup steps	1000	Integer
policy.scheduler_decay_steps	Learning rate decay steps	10000	Integer
policy.scheduler_decay_lr	Final learning rate after decay	0.0	Float

Verification

Training Monitoring

The training details page parses logs per framework and plots the metric curves. Metrics for the jax framework:

Metric	Meaning	Criteria
step	Number of completed training steps	Increases step by step up to the configured steps
loss	Training loss	Decreases overall; if it stops decreasing for a long time, check the learning rate and data
gradient_norm	Gradient norm	Does not keep growing
param_norm	Parameter norm	Changes slowly over training

Metrics for the lerobot framework:

Metric	Meaning	Criteria
step	Number of completed training steps	Increases step by step up to the configured steps
sample	Number of processed samples	Increases with step
episode	Number of processed episodes	Increases with step
epoch	Number of trained epochs	Increases with step
loss	Training loss	Decreases overall
gradient_norm	Gradient norm	Does not keep growing
learning_rate	Current learning rate	Changes according to the schedule
update_time_s	Time per update step	Stable, does not keep growing
data_time_s	Data loading time	Stable

Verification Checklist

The training service determines the final task state from the container exit code: exit code 0 is recorded as Success, and a nonzero code as Failed.

Check	Pass criteria
Task status	The training details page shows Success and the container exit code is 0
Output directory	The jax framework writes to docker_train/train/ under the output directory; the lerobot framework writes to checkpoints/
Checkpoints	Checkpoints lists the checkpoints; jax contents include params/, and lerobot contents include pretrained_model/; the entry that last points to has type Final, and the rest have type Step
Metric curves	Training Metrics plots step, loss, and gradient_norm for jax; for lerobot it also includes sample, episode, epoch, and learning_rate
Logs	Real-time Logs can be read without repeated errors

Checkpoints listed in Checkpoints can be downloaded, synced to object storage, or used to create an inference service; see Model Inference for the inference side.

Error Handling

Symptom	Possible cause	Action	Owner
Training fails immediately with a dataset-not-found error	The dataset lacks meta/info.json, or the selected source path is wrong	Select the correct Export Records entry or uploaded dataset again	Project Manager
Out of GPU memory	batch_size is too large, or full fine-tuning runs on a single GPU	Lower batch_size; keep LoRA enabled on a single GPU	Algorithm Engineer
Invalid FSDP device count error	fsdp_devices exceeds the visible GPU count or is not a divisor of it	Set fsdp_devices to auto, or to a divisor of the visible GPU count	Algorithm Engineer
Save interval greater than training steps error	save_interval or save_freq exceeds steps	Reduce the save interval to no more than steps	Algorithm Engineer
Single-GPU and multi-GPU results differ greatly	A single GPU defaults to LoRA and multiple GPUs default to FSDP, so the number of trained parameters differs	Fix the framework, fsdp_devices, batch_size, and random seed, then compare again	Algorithm Engineer
Loss stops decreasing for a long time	Unsuitable learning rate, missing task text, or insufficient data coverage	Check prompt and the dataset; adjust learning_rate	Algorithm Engineer
Training location unavailable	The platform has not enabled that training location, or its GPU is unavailable	Switch to another training location; ask an administrator to check the Training Service	Administrator

Related Pages

Page	Purpose
Model Training	Creating, monitoring, and quotas for training tasks
LeRobot datasets and training	Dataset export and training overview
LeRobot v2 and v3 format differences	Differences between the two data formats and migration
Data export	Generate a LeRobot training package