EvoStudio HIL Runtime v0.2.3
Reproducible installation bundle for EvoStudio policy_hil_continuous on the
PiPER-X collection hosts. Redistributable Linux/Python 3.12 dependencies are
bundled in a wheelhouse so collection hosts do not depend on PyPI throughput.
This repository intentionally does not redistribute the NVIDIA CUDA wheels;
the installer resolves only those exact pinned versions from the package index.
The BSD/MIT-licensed PyTorch, torchvision, torchcodec and Triton wheels are part
of the checksummed wheelhouse.
Validated reference environment:
- Ubuntu 22.04 x86_64
- Python 3.12.13
- NVIDIA driver 580.173.02
- PyTorch 2.11.0+cu128
- torchvision 0.26.0+cu128
- LeRobot 0.6.0
- EvoStudio inference SDK 0.7.0b7
- EvoStudio HIL runtime 0.2.3
The bundled wheels are immutable and covered by SHA256SUMS. Checkpoints,
device identity, camera/CAN serials, calibration, Cloud credentials and datasets
are never part of this repository.
Install
Download a pinned Hub revision, then install it atomically:
HF_ENDPOINT=https://hf-mirror.com hf download Travor278/evostudio-hil-runtime \
--revision runtime-v0.2.3-r2 \
--local-dir /tmp/evostudio-hil-runtime-v0.2.3
sudo bash /tmp/evostudio-hil-runtime-v0.2.3/install.sh \
--python /home/zhaobo/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/bin/python3.12 \
--destination /opt/evostudio-hil-runtime
sudo bash /tmp/evostudio-hil-runtime-v0.2.3/install-rgbd.sh \
--destination /opt/evostudio-rgbd
The installer keeps any previous destination as a timestamped backup. It does not copy or modify host-specific EvoStudio configuration.
Redistribution boundary
The repository contains the redistributable project and upstream wheels required for the validated runtime. NVIDIA proprietary dependencies are downloaded from their official package source during installation and remain subject to NVIDIA's license terms.
flash-attn is intentionally excluded: the validated Pi0.5 checkpoint uses the
stock PyTorch attention path and neither its configuration nor the native RTC
loader selects FlashAttention. Excluding the unused binary saves about 1 GB of
installed size without changing RTC outputs.