AXIS Pi0.5 baseline
An inference-only Pi0.5 checkpoint fine-tuned on rendered AXIS training replays.
Weights are native OpenPI JAX / Orbax OCDBT, with params/ at the repository root.
This project-hosted release preserves the model weights and normalization files;
it does not represent a new training run.
Input and output
Use the public OpenRoboto evaluation harness
and its pi05_axis_joint configuration. Inputs are the scene's camera0 RGB image,
native 9D joint state and task instruction. Outputs are absolute 9D joint targets,
not LIBERO 7D relative EEF actions. Use the supplied normalization statistics at
assets/axis-v0.1-task501-runtime-v1/norm_stats.json.
The AXIS guide documents installation, image/state transforms and training replay rendering with OSMesa. Cache scene assets before attempting an offline evaluation.
uv run python libero_eval/run_eval.py \
--model openroboto-ai/pi05-axis-baseline \
--commit-id 55f8b28ed021f7ee0bef02cde114a7b5dcae9d5c \
--benchmark axis_v1.0 --backbone pi0.5 \
--num-trials 20 --gpus 0
The model revision above pins the verified inference files. Pin the evaluator
to 2f69d117517f8b388d2d01a94964df63f1b6620e before running this command.
Historical reference result
evaluation_summary.json records 448 successes out of 600 episodes (74.67%) on
the earlier axis_v0.2 protocol: 30 tasks, 20 trials each, fixed scenes, policy
seed 20260907, 120-step limit, replanning every 10 controls and 5 Hz control.
Evaluation used training-task scenes; this is not a held-out generalization score
or a new competition-6 evaluation. Internal training-path metadata and optimizer
state are not distributed. They are not required by the released inference contract.
Terms
The weights have been modified from upstream Pi0.5 through AXIS fine-tuning.
The accompanying LICENSE_GEMMA.txt, LICENSE_OPENPI.txt and NOTICE are retained.
Gemma Terms of Use, including Section 3.2 restrictions, apply to this model and
its derivatives. CHECKSUMS.sha256 lists the distributed inference and license files.