Skip to content

Latest commit

 

History

History
338 lines (253 loc) · 16.4 KB

File metadata and controls

338 lines (253 loc) · 16.4 KB

π₀.₅ (Pi05) Policy

π₀.₅ is a Vision-Language-Action model with open-world generalization, from Physical Intelligence. The LeRobot implementation is adapted from their open source OpenPI repository.

Model Overview

π₀.₅ represents a significant evolution from π₀, developed by Physical Intelligence to address a big challenge in robotics: open-world generalization. While robots can perform impressive tasks in controlled environments, π₀.₅ is designed to generalize to entirely new environments and situations that were never seen during training.

The Generalization Challenge

As Physical Intelligence explains, the fundamental challenge isn't performing tasks of agility or dexterity, but generalization, the ability to correctly perform tasks in new settings with new objects. Consider a robot cleaning different homes: each home has different objects in different places. Generalization must occur at multiple levels:

  • Physical Level: Understanding how to pick up a spoon (by the handle) or plate (by the edge), even with unseen objects in cluttered environments
  • Semantic Level: Understanding task semantics, where to put clothes and shoes (laundry hamper, not on the bed), and what tools are appropriate for cleaning spills
  • Environmental Level: Adapting to "messy" real-world environments like homes, grocery stores, offices, and hospitals

Co-Training on Heterogeneous Data

The breakthrough innovation in π₀.₅ is co-training on heterogeneous data sources. The model learns from:

  1. Multimodal Web Data: Image captioning, visual question answering, object detection
  2. Verbal Instructions: Humans coaching robots through complex tasks step-by-step
  3. Subtask Commands: High-level semantic behavior labels (e.g., "pick up the pillow" for an unmade bed)
  4. Cross-Embodiment Robot Data: Data from various robot platforms with different capabilities
  5. Multi-Environment Data: Static robots deployed across many different homes
  6. Mobile Manipulation Data: ~400 hours of mobile robot demonstrations

This diverse training mixture creates a "curriculum" that enables generalization across physical, visual, and semantic levels simultaneously.

Installation Requirements

  1. Install LeRobot by following our Installation Guide.

  2. Install Pi0.5 dependencies by running:

    pip install -e ".[pi]"

    If you installed LeRobot from PyPI:

    pip install 'lerobot[pi]'

Usage

To use π₀.₅ in your LeRobot configuration, specify the policy type as:

policy.type=pi05

Training

Quickstart on LIBERO

Finetune the LIBERO base model on lerobot/libero, a ~1.9 GB video-encoded copy of the demonstrations behind the results below.

It carries the keys π₀.₅ reads, which are also the ones the LIBERO environment produces at evaluation time:

Feature Shape in the dataset How π₀.₅ consumes it
observation.images.image 256×256×3, agentview resized to 224×224
observation.images.image2 256×256×3, wrist resized to 224×224
observation.state 8 discretized into 256 bins and written into the prompt
action 7 padded to 32 internally; the loss uses the first 7 dims

No --rename_map is needed here — the keys already match; see Rename Map and Empty Cameras if yours differ.

π₀.₅ uses the gated [google/paligemma-3b-pt-224](https://huggingface.co/google/paligemma-3b-pt-224) tokenizer — accept its license on the Hub and log in with `hf auth login` before training.

Sized for a single 80 GB GPU:

lerobot-train \
    --dataset.repo_id=lerobot/libero \
    --policy.type=pi05 \
    --policy.pretrained_path=lerobot/pi05_libero_base \
    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
    --policy.n_action_steps=10 \
    --policy.empty_cameras=1 \
    --policy.freeze_vision_encoder=false \
    --policy.train_expert_only=false \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --policy.device=cuda \
    --policy.push_to_hub=false \
    --output_dir=./outputs/pi05_libero \
    --job_name=pi05_libero \
    --batch_size=64 \
    --num_workers=8 \
    --steps=30000 \
    --save_freq=5000 \
    --seed=1000

Mean/std normalization, not π₀.₅'s quantile default — matching pi05_libero_finetuned_v044, the checkpoint the results below were measured on.

--policy.n_action_steps=10 and --policy.empty_cameras=1 are explicit because --policy.pretrained_path loads weights only — lerobot/pi05_libero_base stores both, and they would otherwise fall back to 50 and 0 (see Loading a checkpoint).

Then evaluate a checkpoint with lerobot-eval and compare against the reference success rates — see LIBERO.

Quantile statistics

π₀.₅ normalizes STATE and ACTION with quantiles, so your dataset's meta/stats.json needs q01 and q99. Older datasets carry only min/max/mean/std and fail on the first batch:

ValueError: QUANTILES normalization mode requires q01 and q99 stats

Recompute them:

lerobot-edit-dataset \
    --repo_id your_dataset \
    --new_repo_id your_dataset \
    --operation.type recompute_stats \
    --operation.overwrite true

The result lands in $HF_LEROBOT_HOME/your_dataset, not the cache --dataset.repo_id reads — so train with --dataset.root=$HF_LEROBOT_HOME/your_dataset, or add --push_to_hub true above.

Or keep the dataset as-is and pass --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}'.

Recording, resuming, and merging aggregate quantiles from per-episode summaries, so meta/stats.json ends up holding a conservative envelope (min for q <= 50, max for q > 50) rather than whole-dataset quantiles. To estimate the latter, scan every episode with a running histogram:

python src/lerobot/scripts/augment_dataset_quantile_stats.py \
    --repo-id=your_dataset \
    --overwrite \
    --skip-images

--skip-images keeps the existing image statistics and avoids video decoding when only STATE/ACTION need recomputing, and --root reads a local dataset instead of the Hub. These values are histogram estimates, subject to discretization and rebinning error, so they can differ from the conservative ones — which changes π₀.₅'s normalized targets and therefore its loss scale. Statistics already saved inside an existing checkpoint are not affected.

Training Command Example

The same finetune with the VLM frozen: less memory, at some cost in success rate. Swap --dataset.repo_id for your own dataset.

lerobot-train \
    --dataset.repo_id=lerobot/libero \
    --policy.type=pi05 \
    --policy.pretrained_path=lerobot/pi05_libero_base \
    --policy.normalization_mapping='{"ACTION": "MEAN_STD", "STATE": "MEAN_STD", "VISUAL": "IDENTITY"}' \
    --policy.n_action_steps=10 \
    --policy.empty_cameras=1 \
    --policy.freeze_vision_encoder=true \
    --policy.train_expert_only=true \
    --policy.gradient_checkpointing=true \
    --policy.dtype=bfloat16 \
    --policy.device=cuda \
    --policy.push_to_hub=false \
    --output_dir=./outputs/pi05_libero_expert \
    --job_name=pi05_libero_expert \
    --batch_size=64 \
    --num_workers=8 \
    --steps=30000 \
    --save_freq=5000 \
    --seed=1000

Key Training Parameters

  • --policy.compile_model=true: Enables model compilation for faster training
  • --policy.gradient_checkpointing=true: Reduces memory usage significantly during training
  • --policy.dtype=bfloat16: Use mixed precision training for efficiency
  • --batch_size=64: Batch size for training, adapt this based on your GPU memory
  • --policy.pretrained_path=lerobot/pi05_base: The base π₀.₅ model you want to finetune, options are:

Loading a checkpoint

The two forms are not interchangeable:

--policy.path --policy.pretrained_path
Loads weights and the checkpoint's config.json weights only
Feature names from the checkpoint from your dataset, after --rename_map
Stored settings, e.g. n_action_steps inherited reset to the defaults
--policy.type must be omitted required
--rename_map needed when your camera keys differ supported for canonicalizing dataset keys

Use --rename_map with either loading form when raw dataset feature names differ from the policy feature names. For example, the RoboMME fine-tuning command maps its raw image, wrist_image, state, and actions columns while initializing weights with --policy.pretrained_path.

Training Parameters Explained

Parameter Default Description
freeze_vision_encoder false Do not freeze the vision encoder
train_expert_only false Do not freeze the VLM, train all parameters

💡 Tip: Setting train_expert_only=true freezes the VLM and trains only the action expert and projections, allowing finetuning with reduced memory usage.

Short-Horizon Observation Memory (MEM)

Pi05 can optionally use short-horizon visual and proprioceptive context based on MEM. Both paths are disabled by default, so existing checkpoints and training commands retain the single-frame Pi05 behavior.

This implements MEM's short-horizon memory only — the video encoder of section III-C and the proprioceptive projection of section III-D. MEM's long-horizon language memory (section III-B), in which a high-level policy predicts the next subtask and a compressed natural-language summary of what has happened so far, is not implemented. In the paper's ablations, video memory alone recovers only part of full MEM's task progress on long-horizon tasks.

Introducing memory when finetuning from a checkpoint that was pretrained without it also corresponds to the paper's weaker MEM-Posttrain-Only ablation. MEM's headline results come from pretraining the video encoder on a diverse mixture of robot and non-robot video, which no public Pi05 checkpoint currently provides.

Enable it while finetuning with:

lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
    --policy.pretrained_path=lerobot/pi05_base \
    --policy.use_visual_memory=true \
    --policy.memory_frames=6 \
    --policy.memory_stride=30 \
    --policy.memory_temporal_attention_every=4 \
    ...

memory_frames includes the current observation. memory_stride is measured in dataset frames, not seconds, so scale it with your dataset's fps: MEM pretrains on six observations one second apart, which is memory_stride=30 for a 30 fps dataset (the default) but memory_stride=10 for a 10 fps dataset such as lerobot/robomme.

Every fourth SigLIP layer replaces its attention with MEM's composed space-time attention: causal temporal attention between matching patch tokens, composed with the standard spatial attention and reusing the pretrained projections, so no learnable parameters are added to the vision tower. Past-frame tokens are dropped once the last such layer has run, keeping the downstream language/action prefix length unchanged.

Unavailable history at the beginning of an episode is padded and masked. At inference, Pi05 maintains the same strided image history internally and clears it whenever the policy is reset.

Historical proprioception is an independent option:

--policy.use_proprioceptive_memory=true

When enabled, one projected continuous state token is added for each retained frame, and the discretized state is removed from the text prompt so the state is represented exactly once (MEM section III-D). This changes the prompt format, so enable it from the start of training rather than partway through. Visual and proprioceptive memory use the same memory_frames and memory_stride settings, but either path can be enabled independently. Train or finetune with the selected memory options; enabling them only at inference does not give a single-frame checkpoint learned memory behavior.

Under PEFT, proprio_history_proj is trained and saved in full via modules_to_save, since it has no pretrained weights to adapt.

At inference, MEM keeps one batched history queue and assumes every batch row shares episode boundaries. lerobot-eval satisfies this by calling policy.reset() before each batched rollout and before resetting the vector environment. Independently autoresetting rows in an asynchronous vector batch is not supported; start a new rollout so the complete policy and action queues are reset together.

Relative Actions

By default, π₀.₅ predicts absolute actions. You can enable relative actions so the model predicts offsets relative to the current robot state. This can improve training stability for certain setups.

To use relative actions, first recompute your dataset stats in relative space via the CLI:

lerobot-edit-dataset \
    --repo_id your_dataset \
    --operation.type recompute_stats \
    --operation.relative_action true \
    --operation.chunk_size 50 \
    --operation.relative_exclude_joints "['gripper']" \
    --push_to_hub true

Or equivalently in Python:

from lerobot.datasets import LeRobotDataset, recompute_stats

dataset = LeRobotDataset("your_dataset")
recompute_stats(dataset, relative_action=True, chunk_size=50, relative_exclude_joints=["gripper"])
dataset.push_to_hub()

The chunk_size should match your policy's chunk_size (default 50 for π₀.₅). relative_exclude_joints lists joint names that should remain in absolute space (e.g. gripper commands). Use --push_to_hub true to upload the updated stats to the Hub.

Then train with relative actions enabled:

lerobot-train \
    --dataset.repo_id=your_dataset \
    --policy.type=pi05 \
    --policy.use_relative_actions=true \
    --policy.relative_exclude_joints='["gripper"]' \
    ...

Performance Results

Libero Benchmark Results

π₀.₅ has demonstrated strong performance on the Libero benchmark suite. To compare and test its LeRobot implementation, we finetuned the libero base model for an additional 6k steps on the Libero dataset and compared the results to the OpenPI reference results.

Benchmark LeRobot Implementation OpenPI Reference
Libero Spatial 97.0% 98.8%
Libero Object 99.0% 98.2%
Libero Goal 98.0% 98.0%
Libero 10 96.0% 92.4%
Average 97.5% 96.85%

These results demonstrate π₀.₅'s strong generalization capabilities across diverse robotic manipulation tasks. To reproduce these results, you can follow the instructions in the Libero section.

License

This model follows the Apache 2.0 License, consistent with the original OpenPI repository.