pi05-agilex3-2bit
A 2-bit quantized version of π0.5 for the AgileX Piper robotic arm
3 min readupdated
The pi05-agilex3-2bit model by ForeverBlue is a 2-bit quantized (weight-compressed) version of agilex3_pi05 (a full-parameter fine-tune of π0.5 on three manipulation tasks, step 5999 EMA) built on the openpi library. It is designed for testing on a real AgileX Piper robot and benchmarking accuracy against the full-precision variant: input/output interfaces, commands, and the 10 Hz control frequency are identical to the original. Quantization was done via calibration only (on 256 demonstration observations with lighting and crop augmentations) with no additional fine-tuning, and the weights are saved unpacked in standard PyTorch format (model.safetensors) to run in a default openpi environment without code changes.
Key features
- Aggressive 2-bit quantization: all 414 linear layers across the vision encoder, language model, and action expert (2.7B parameters) are compressed to an average of 2.016 bits per weight, while the embedding table is cast to INT8.
- Full openpi compatibility: the model uses the same normalization statistics in
norm_stats.jsonand plugs into existingagilex3_pi05.pyscripts simply by changing the--checkpointpath. - Close to the original: the prediction deviation of the 2-bit version from the full-precision model is only 1.2x the natural variance of the original model itself when changing the initial noise seed.
- Two integration modes: supports both calling the policy directly from the robot's Python code and running it over the network via a local policy server.
What you can do
- Run A/B tests on a real AgileX Piper: alternate between the full-precision and 2-bit models across three tasks (
cluster_grasp,cabinet_lemon,circle) using identical object layouts to measure real-world success rates. - Collect data for future calibration: record dual-camera streams, robot states, and commands while running the 2-bit model to improve subsequent quantization iterations.
Example
import os; os.environ.setdefault("JAX_PLATFORMS", "cpu")
import sys; sys.path.insert(0, "/path/to/agilex3_pi05")
from agilex3_pi05 import AgileXPi05Policy
policy = AgileXPi05Policy(checkpoint="/path/to/pi05-agilex3-2bit/checkpoint")
actions = policy.predict(fixed, wrist, state, instruction) # (10, 7)
Read more
DetailsOffline tests (256 held-out observations): mean joint error increased from 0.0168 rad in the full-precision model to 0.0237 rad (17% of the stationary baseline error), gripper error went from 2.31 mm to 3.16 mm, and max single-step deviation was 0.217 rad.
The
model.safetensors file size is 14.5 GB (unpacked weights for PyTorch compatibility; actual 2-bit layer footprint is 0.68 GB vs. 5.41 GB in BF16, though a custom runtime is not included in the repository).Before running, you need to apply the
transformers_replace patch for transformers 4.53.2 and set JAX_PLATFORMS=cpu.The first
predict call triggers torch.compile compilation (around 472 seconds on an RTX 6000 Ada), after which median inference time is 0.067 s.









Comments