Quick Start: Pi0.5 Training#

This guide walks you through launching Pi0.5 (π₀.₅) fine-tuning and offline evaluation in the LoongForge framework.

0. Resource Preparation#

0.1 Model Weights#

The Pi0.5 main weights are specified by CHECKPOINT_PATH and loaded via --pretrained-checkpoint. The original HuggingFace safetensors format is supported:

hf download lerobot/pi05_base --local-dir /workspace/huggingface.co/lerobot/pi05_base

0.2 Tokenizer / Processor#

Pi0.5 requires a tokenizer, specified by TOKENIZER_PATH and loaded via --tokenizer-path:

huggingface-cli download google/paligemma-3b-pt-224 --local-dir /workspace/huggingface.co/google/paligemma-3b-pt-224

0.3 Dataset#

The script uses LIBERO data in LeRobot v3.0 format by default. If the dataset comes from a HuggingFace Dataset, download it to the same directory:

hf download lerobot/libero_10 --repo-type dataset --local-dir /workspace/libero_10

1. Data Configuration#

For the standard LeRobot v3.0 format, no extra offline preprocessing is required. The GR00T sample transform and batch collator run online during training.

Default data configuration:

data:
  image_size: 224
  image_normalize_mode: identity

2. Launch Training#

First, set the paths:

cd /workspace/LoongForge

export LOONGFORGE_PATH=/workspace/LoongForge
export TOKENIZER_PATH=/workspace/huggingface.co/google/paligemma-3b-pt-224

export CHECKPOINT_PATH=/workspace/workspace/huggingface.co/lerobot/pi05_base

export DATA_PATH=/workspace/libero_10
export OUTPUT_DIR=/workspace/outputs/pi05
export TENSORBOARD_PATH=/workspace/tensorboard-log/pi05

2.1 Script#

The DDP script already enables the main performance optimizations:

DDP:

bash examples/embodied/pi05/run_pi05_ddp_finetune.sh

DDP + ZeRO-1:

bash examples/embodied/pi05/run_pi05_ddp_zero1_finetune.sh

FSDP:

bash examples/embodied/pi05/run_pi05_fsdp_finetune.sh

2.2 Correctness Verification#

To ensure training accuracy is not affected by the optimizations, we ran a step-by-step action loss comparison between LoongForge’s Pi0.5 implementation and the official one under identical data, weights, and training configuration. The results show that LoongForge’s performance optimizations are lossless with respect to training accuracy: alt text

3. Evaluation#

LoongForge provides a standalone benchmark evaluation module (LIBERO / CALVIN / SimplerEnv / RoboTwin / ManiSkill). The benchmark client and policy server are decoupled via WebSocket.

3.1 Quick Run LIBERO#

Edit the evaluation YAML

(Example: examples/embodied/pi05/eval/configs/libero/smoke_steps10.yaml), and fill in:

* `server.ckpt_path`: the checkpoint from training (a directory containing `model.safetensors` or the weight file)
* `server.dataset_statistics_path`: dataset statistics (for action denormalization; Pi0.5 default is q99)
* `server.tokenizer_path`: PaliGemma tokenizer

Launch:

cd $LOONGFORGE_PATH
# Uses smoke_steps10.yaml by default; can be overridden via CONFIG
CONFIG=examples/embodied/pi05/eval/configs/libero/smoke_steps10.yaml \
  bash examples/embodied/pi05/eval/run_libero_eval.sh

Quick Start: Pi0.5 Training

Contents