Machines and Network

Machines & Network

• NVIDIA Linux workstation (gpu = 192.168.11.130) — RTX 4090; GPU training + eval. The workhorse.
• MacBook Pro (M2 Pro) — development + sim teleop. No CUDA.
• Robot PC (Intel, robotdisco = 192.168.11.80) — hardware teleop/recording: arm + hand + cameras + SpaceMouse/gamepad.
• Robot subnet 192.168.11.x — UR5e .21 (RTDE :30004, dashboard :29999), AmazingHand ESP32 .117 (TCP :8765). The subnet is isolated (no internet); install on the robot PC by tunnelling pip over SSH through a machine that has internet.

Envs: lerobot 0.4.4 in the project .venv on both the GPU box and the robot PC (.80, Python 3.10) — so trained checkpoints load across them without a version mismatch; the Mac dev env is lerobot 0.5.x / Python 3.12 (sim + development only, not in the train→deploy path).

Robot PC (.80): run.sh

Every robot-PC command goes through ~/VLA/run.sh, which cds to ~/VLA, sets PYTHONPATH=., PYTHONNOUSERSITE=1 (isolates from a conflicting ~/.local ML stack), and HF_HUB_OFFLINE=1 (the subnet has no internet — stops lerobot/HF from hanging on DNS), and uses the project venv:

~/VLA/run.sh -m ur5e_lerobot.teleop.manual_panel --engine hardware