pi05

Physical Intelligence's pi0.5 base checkpoint (pi05_base, openpi) behind the arena_planners bridge, as a zero-shot navigation baseline. It drives the robot through the two action dims that command a mobile manipulator's base, [v, omega], from the front camera and a text instruction, the way VEGA (arXiv 2606.18426, sec. 5.1) evaluates it. No public navigation fine-tune of pi0.5 exists, so this is the base model as released.

arena feature planners add pi05
planners/pi05/server/setup.sh
arena launch sim:=gazebo robot:=jackal[camera=vln_navid] robot.planner:=pi05

robot.planner:=pi05 selects robot.mobile:=vla. The planner accepts the instruction only (goal_inputs: [instruction]), the goal pose never reaches it and decides success alone. It has no stop head and sends no signals, Arena judges arrival.

Model inputs and outputs

openpi's pi05_aloha inference config (openpi's own default serve_policy.py --env ALOHA, which also serves pi05_base) with two changes in server/serve.py: the Mobile ALOHA norm stats (trossen_mobile) instead of the static ALOHA ones, and an output transform that keeps the 16 action dims of that embodiment instead of the 14 arm dims (AlohaOutputs drops the base).

Input	Value
base_0_rgb	the robot camera frame, uint8 RGB at its native size, resize_with_pad to 224x224 on the server
left_wrist_0_rgb, right_wrist_0_rgb	black, image mask false
state (14 dims)	the embodiment's mean state from its norm_stats.json, constant: the robot has no arms
prompt	the goal instruction, whitespace-stripped

The model samples 50 actions of 32 dims with 10 flow-matching steps, unnormalized with quantile stats ([-1, 1] maps to [q01, q99]). Dims 14 and 15 are the base's linear and angular velocity, Mobile ALOHA's base_action. For trossen_mobile, q01/q99 are -0.219/0.398 m/s and -0.314/0.252 rad/s.

The chunk is 50 Hz (1 s). The planner averages it into 0.1 s twists (10 entries) and returns a velocity_chunk at rate_hz: 10, so every entry plays for one bridge tick. The bridge runs 0.5 s of it (chunk.execute_s, pi0.5's replanning cadence of 25 steps), clamps each twist to the robot's velocity limits, and sends the next frame. There is no rescaling between Mobile ALOHA's base and the robot's.

robot.mobile.instruction:=<text> stands in for goals that carry none. Without an instruction or a camera frame the planner stands still without calling the model.

Server

The model runs in a separate process: openpi pins Python 3.11, numpy<2, torch 2.7.1 and transformers 4.53.2, which cannot share the planner venv. server/ is a uv project for that venv, server/serve.py builds the policy and serves it with openpi's websocket server on 127.0.0.1. The planner spawns it at init, waits for its metadata frame, talks openpi's msgpack wire format to it, and stops it at exit.

server/setup.sh (idempotent) syncs that venv (torch from the cu128 index, which carries Blackwell sm_120 kernels), applies openpi's transformers_replace patches into the venv's own copy of transformers (UV_LINK_MODE=copy, so the uv cache stays clean), and converts the JAX checkpoint to PyTorch with openpi's examples/convert_jax_model_to_pytorch.py at the pinned commit, copying the norm-stat assets the converter misses.

Config

Set in planner.yaml > config, override per launch with robot.mobile.config.<key>:=<value>.

Key	Default	Effect
checkpoint	model/pi05_base_pytorch	Checkpoint dir. A dir with model.safetensors loads in PyTorch (JAX pinned to CPU), model/pi05_base loads the original JAX checkpoint.
embodiment	trossen_mobile	Norm stats and state: trossen_mobile (Mobile ALOHA) or arx_mobile, both with the base at dims 14 and 15.
device	auto	PyTorch device, auto picks CUDA when available.
flow_steps	10	Flow-matching integration steps.
compile_mode	empty	torch.compile mode for sample_actions (openpi's default is max-autotune), empty compiles nothing.
server_python	server/.venv/bin/python	Interpreter of the server venv.
tokenizer	model/paligemma_tokenizer.model	PaliGemma tokenizer, copied into openpi_data so the server never downloads it.
openpi_data	model/openpi-data	OPENPI_DATA_HOME of the server.
port	0	Server port on 127.0.0.1, 0 picks a free one.
startup_timeout_s	900	Time allowed for the server to load the checkpoint.

Weights and license

weights.yaml fetches the 12.4 GB gs://openpi-assets/checkpoints/pi05_base (params plus the trossen_mobile and arx_mobile norm stats) and the PaliGemma tokenizer from gs://big_vision. The pi0.5 weights derive from PaliGemma and are subject to the Gemma Terms of Use, see openpi's LICENSE_GEMMA.txt. The code is Apache-2.0, see LICENSE and ATTRIBUTION.md.

@article{intelligence2025pi05,
  title={$\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization},
  author={{Physical Intelligence} and others},
  journal={arXiv preprint arXiv:2504.16054},
  year={2025}
}