G1 pi0.5 deployment bundle

Migration snapshot of an on-Orin pi0.5 inference stack for a Unitree G1: a WebSocket policy server running pi0.5-ggml (CUDA, sm_87) on the robot's Jetson Orin NX, plus the GGUF exports it serves and the client it talks to.

Captured 2026-09-17 from unitree-g1-nx (Jetson Orin NX, JetPack / CUDA 12.6, 15 GB unified memory).

Contents

path	size	what
g1-migration-code.zip	124 MB	all source (see below)
models/gguf_fp16/	6.3 GB	FP16 GGUF export — the accuracy baseline
models/gguf_q8_h32/	3.9 GB	Q8_0 GGUF export — the quantized candidate
models/stack-cube-eef-24k_pytorch/	7.3 GB	source PyTorch checkpoint (copy of LGG100/stack-cube-eef-24k)

g1-migration-code.zip unpacks to:

pi0.5-server/    the policy server + open-loop eval harness (written for this deployment)
g1-client/       robot-side client        (git: luo-jingw/g1-client)
pi05-main/       pi0.5-ggml, master       (git: luo-jingw/pi0.5-ggml)
pi05-neowise/    pi0.5-ggml, amir_neowise_q8 branch
datasets/        3 episodes of LGG100/Stack-the-cubes, for the open-loop test
models/ckpt_h32/config.json   the corrected export config (see below)

build/ directories, .git/, and __pycache__ were excluded. The three source trees are live git repos upstream — prefer cloning them and using this zip only for pi0.5-server/ and to see exactly which commits were deployed.

Start at pi0.5-server/README.md.

Two corrections baked into these exports

Both were real bugs found while bringing this up; the models here have them fixed.

1. action_horizon must be 32, not 10. The PyTorch checkpoint's config.json says action_horizon = 10, which is a placeholder borrowed during conversion, not the training value. Exporting with 10 made the arm travel only 4–5 mm per chunk — it looked like the robot was barely moving. At 32 it is 20–38 mm. Corroborated independently: GGUFs built from this checkpoint through the correct pipeline elsewhere all carry action_horizon = 32.

models/ckpt_h32/config.json in the zip is the corrected config used for these exports.

2. Do not feed the model a discrete state vector. pi05_g1_eef was not trained with state as an input (confirmed by the model owner). pi0.5-ggml only wants one for checkpoints trained with discrete_state_input=True (include/pi05.h:92); passing it otherwise wraps the prompt into a Task:/State: form the model never saw. The server accepts observation.state on the wire but does not forward it unless --discrete-state is given.

Measured open-loop, FP16, mean EEF position error over a chunk: 22.2 mm with state vs 9.0 mm without.

Measured on this hardware

Open-loop against LGG100/Stack-the-cubes (24 samples / 3 episodes, --steps 10, state omitted). Mean EEF position error, both arms:

	t=0	whole chunk	latency	peak host mem
FP16 (stock ggml)	8.7 mm	9.0 mm	~1484 ms	9.8 GB
Q8_0 + NEOWISE variant 4	24.4 mm	28.5 mm	~1150 ms	10.0 GB
"hold current pose" baseline	—	17.0 mm	—	—

The NEOWISE Q8 INT8 kernel (amir_neowise_q8 branch) costs ~3.2× accuracy here and does not beat a do-nothing baseline. Its published −25 % speedup was measured on AGX Thor (Blackwell); docs/neowise.md listed Orin as unmeasured. On this Orin NX it is also slower than stock in end-to-end benchmark (1227.8 ms vs 1199.7 ms) and uses 2.5 GB more memory.

Caveats on the accuracy numbers: Stack-the-cubes stores joint actions while this is an EEF checkpoint, so ground truth is pushed through the same forward kinematics the runtime uses — this is not the checkpoint's training set, and the FK frame convention is unverified against training. A frame mismatch would shift all absolute errors; the FP16-vs-Q8 comparison is unaffected. The dataset is also Unitree_G1_Dex1_Sim (simulation), and open-loop error is not a success rate.

Rebuilding

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \
    -DGGML_CUDA=ON -DGGML_CUDA_NO_VMM=ON \
    -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
    -DCUDAToolkit_ROOT=/usr/local/cuda \
    -DCMAKE_CUDA_ARCHITECTURES=87 \
    -DCMAKE_CUDA_RUNTIME_LIBRARY=Shared \
    -DPython_EXECUTABLE=<py3.11 env>/bin/python
cmake --build build -j6

sm_87 for Orin; CMake < 3.24 cannot auto-detect it. nvcc is not on PATH on JetPack. Python 3.11 + pybind11. For the neowise build add -DUSE_NEOWISE_Q8=ON -DCUSTOM_KERNEL_VERSION=4 (variant 6 is a Blackwell CUTLASS kernel and cannot run on Ampere).