π₀.₅ for RoboTwin 2.0

This is a LeRobot π₀.₅ Vision-Language-Action policy fine-tuned from lerobot/pi05_base on the lerobot/robotwin_unified dataset. It predicts joint-space action chunks for the 14-DoF Aloha-AgileX bimanual robot used by RoboTwin 2.0.

The checkpoint in this repository is the final checkpoint from training step 50,000. The model consumes language instructions, one 14-dimensional robot state, and three RGB camera views.

Model details

Property	Value
Policy	π₀.₅ (`pi05`)
Fine-tuned from	`lerobot/pi05_base`
Framework	LeRobot / PyTorch
Robot	Aloha-AgileX bimanual, 14 DoF
Action representation	Absolute joint-space actions
Action chunk size	50
Actions executed per prediction	50
Inference steps	10
Internal image resolution	224 × 224
Model dtype	bfloat16
License	Apache-2.0

Inputs and outputs

The feature names below are part of the checkpoint configuration. Camera names must match these keys, or be mapped to them with --rename_map.

Inputs

Feature	Type	Shape
`observation.state`	State	`(14,)`
`observation.images.cam_high`	RGB image	`(480, 640, 3)`
`observation.images.cam_left_wrist`	RGB image	`(480, 640, 3)`
`observation.images.cam_right_wrist`	RGB image	`(480, 640, 3)`
`task`	Natural-language instruction	string

Output

Feature	Type	Shape
`action`	Absolute joint-space action	`(14,)` per step, 50-step chunks

The 14 state/action dimensions are ordered as follows:

left_waist, left_shoulder, left_elbow, left_forearm_roll,
left_wrist_angle, left_wrist_rotate, left_gripper,
right_waist, right_shoulder, right_elbow, right_forearm_roll,
right_wrist_angle, right_wrist_rotate, right_gripper

Training data

The policy was trained on lerobot/robotwin_unified in LeRobot v3.0 format. The dataset metadata available for this training run reports:

27,500 episodes
6,075,103 frames
30 FPS
three RGB views: high, left wrist, and right wrist
14-dimensional Aloha joint state and action

RoboTwin 2.0 covers 50 bimanual manipulation tasks with varied objects, layouts, lighting, backgrounds, and language instructions.

Training configuration

Setting	Value
Training steps	50,000
Batch size	16
Optimizer	AdamW
Peak learning rate	2.5e-5
Weight decay	0.01
LR schedule	Cosine decay with 1,000 warmup steps
Decay steps / final LR	30,000 / 2.5e-6
Seed	1000
Gradient checkpointing	Enabled
`torch.compile`	Enabled (`max-autotune`)
Vision encoder frozen	No
Expert-only training	No
Image augmentation	Disabled
State/action normalization	Mean and standard deviation

The saved preprocessor and postprocessor files contain the normalization state needed for inference; upload them together with model.safetensors and config.json.

Installation

Install LeRobot with the π policy dependencies:

pip install "lerobot[pi]"

RoboTwin evaluation additionally requires Linux, an NVIDIA GPU, and the RoboTwin SAPIEN/CuRobo environment. See the LeRobot RoboTwin guide for simulator setup.

Loading the policy

The checkpoint includes serialized pre- and postprocessing pipelines. Load all three components from the same Hub repository:

import torch

from lerobot.policies import make_pre_post_processors
from lerobot.policies.pi05 import PI05Policy

model_id = "SidneyXie/pi05_robotwin"
device = "cuda"

policy = PI05Policy.from_pretrained(model_id)
policy.eval()

preprocessor, postprocessor = make_pre_post_processors(
    policy.config,
    pretrained_path=model_id,
    preprocessor_overrides={"device_processor": {"device": device}},
)

At inference time, pass a batch containing the three image features, observation.state, and a natural-language task. Use policy.select_action(preprocessor(batch)), then apply postprocessor to the result before sending it to the robot or simulator.

RoboTwin evaluation

RoboTwin's environment camera keys differ from the names stored in this checkpoint, so the rename map is required. A quick five-episode evaluation on one task is:

lerobot-eval \
  --policy.path=SidneyXie/pi05_robotwin \
  --env.type=robotwin \
  --env.task=beat_block_hammer \
  --eval.batch_size=1 \
  --eval.n_episodes=5 \
  --rename_map='{"observation.images.head_camera":"observation.images.cam_high","observation.images.left_camera":"observation.images.cam_left_wrist","observation.images.right_camera":"observation.images.cam_right_wrist"}' \
  --output_dir=outputs/eval/pi05_robotwin/beat_block_hammer

For an official-style result, evaluate 100 episodes per task and report Easy (demo_clean) and Hard (demo_randomized) settings separately. Consult the RoboTwin leaderboard for the current submission protocol.

Evaluation results

Setting	Task	Episodes	Success rate
Easy	`adjust_bottle`	100	100%
Easy	`beat_block_hammer`	100	93%
Easy	`click_alarmclock`	100	90%
Easy	`click_bell`	100	86%
Easy	`grab_roller`	100	99%
Easy	`lift_pot`	100	57%
Easy	`move_can_pot`	100	48%
Easy	`move_pillbottle_pad`	100	74%
Easy	`move_playingcard_away`	100	96%
Easy	`move_stapler_pad`	100	14%
Easy	`pick_diverse_bottles`	100	59%
Easy	`pick_dual_bottles`	100	69%
Easy	`place_a2b_left`	100	49%
Easy	`place_a2b_right`	100	41%
Easy	`place_container_plate`	100	86%
Easy	`place_fan`	100	50%
Easy	`place_mouse_pad`	100	45%
Easy	`place_object_scale`	100	46%
Easy	`place_object_stand`	100	90%
Easy	`place_phone_stand`	100	66%
Easy	`press_stapler`	100	97%
Easy	`rotate_qrcode`	100	71%
Easy	`stamp_seal`	100	45%
Easy	`turn_switch`	100	55%
Easy	Overall (24 tasks)	2400	67.75%

Intended use and limitations

This checkpoint is intended for research on RoboTwin 2.0 and compatible 14-DoF Aloha-style bimanual setups. It expects the same joint ordering, camera semantics, observation preprocessing, and action convention used during training.

It has not been validated for direct deployment on a physical robot.
Distribution shifts in camera placement, calibration, control frequency, joint scaling, objects, or scene appearance can substantially reduce performance.
The policy can produce unsafe or infeasible actions. Use workspace limits, collision checking, emergency stops, and human supervision on real hardware.
This is a learned policy and does not provide correctness or safety guarantees.

References

@article{intelligence2025pi05,
  title   = {Pi 0.5: a Vision-Language-Action Model with Open-World Generalization},
  author  = {Physical Intelligence and Kevin Black and Noah Brown and others},
  journal = {arXiv preprint arXiv:2504.16054},
  year    = {2025}
}

@misc{cadene2024lerobot,
  title        = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in PyTorch},
  author       = {Cadene, Remi and Alibert, Simon and Soare, Alexander and others},
  year         = {2024},
  howpublished = {\url{https://github.com/huggingface/lerobot}}
}