Hardware Notes

Compute and platform decisions, with the reasoning kept alongside each one — shared infrastructure, not scoped to one project

Referenced by the robot arm work (Stage 1, closed / Stage 2, active) and by other projects using the same home server and ML desktop.

On the Pi 4 (1GB) — is it enough?

For teleoperation and inference, yes. For recording, no — this was tested and corrected mid-project. Don't worry about not getting the 2GB model — even 2GB wouldn't be enough to train imitation-learning/VLA policies at reasonable speed; that kind of training really wants a machine with a GPU. The practical split, confirmed by actual testing rather than assumed upfront:

• Pi 4 (1GB) is fine for: running the Voice/Vision Kit software, driving the arm via LeRobot's control/teleoperation scripts, and running already-trained lightweight policies for inference.
• Recording demonstration datasets should not happen on the Pi 4 (1GB): testing showed lerobot-record hit a real ~4-6Hz ceiling against a 30Hz target when running arm control, camera capture, and real-time video encoding together — a genuine hardware limit, not a tuning problem. See the Mac migration log for the full diagnosis.
• Training (Phase 6) should happen elsewhere: a laptop with a decent GPU, or a rented cloud GPU session. Record data off the Pi (a Mac with MPS acceleration proved effective), transfer the dataset if needed, train off-device, then deploy the resulting model back to the Pi for inference.

This is standard practice in the LeRobot ecosystem — the low-power board is the "body" running the final model, not the machine collecting data or doing the heavy training. The 1GB Pi 4 remains a fine choice for its actual role; recording and training were always going to need more capable hardware.

On the PCIe M.2 HAT (Pi 5) — SSD vs. AI accelerator, researched and decided

Investigated whether an M.2 AI accelerator (e.g. Hailo-8L) could enable on-device SmolVLA inference on the Pi 5, as an alternative to the Mac-tethered setup in the SmolVLA extension plan.

Finding: Hailo's toolchain (Dataflow Compiler, Model Zoo) is built and proven for CNN-based vision models (YOLO, classification, segmentation). Community reports (e.g. attempts to deploy DINOv2/v3, a Vision Transformer) describe the compiler lacking support for certain transformer-internal layers. No evidence found of a successful VLA model deployment on Hailo hardware. SmolVLA's transformer/attention-based architecture falls much closer to the "documented trouble" category (ViT-like) than to Hailo's proven CNN use case.

Decision: skip the AI accelerator for this purpose — the engineering risk (undocumented, likely-unsupported operators) isn't worth it for this project's scale. Use the M.2 slot for an NVMe SSD instead — safe, immediately useful for dataset/model storage and swap headroom on the Pi 5's 4GB RAM, no research risk. Revisit the accelerator question only if Hailo's transformer support meaningfully matures, or for a separate, genuinely CNN-shaped task (e.g. real object detection for Phase 4).

New PC ordered — CSL Sprint 5753 (custom config)

Ordered to unblock SmolVLA training (and general heavier local compute going forward) — the Mac's GPU/memory setup wasn't viable for training a VLA-scale model at reasonable speed.

Configuration ordered (CSL Sprint 5753, custom build), €1,818.50:

Reasoning behind the config:

• VRAM range established empirically first, not guessed: ran test workloads on RunPod against rented A4000 and RTX 3090 instances to find the actual working range for this task, before pricing any hardware. Settled on a 16–24 GB VRAM window as the requirement.
• GPU sourcing: first choice was a used RTX 3090 (sits in the target VRAM range) — but used units were running ~€2500, out of budget. Pivoted to a new system with a reasonable total price under €2000, landing on the RTX 5060 Ti 16GB as the in-budget option that still clears the empirically-set VRAM floor.
• Local box vs. cloud split, decided deliberately, not a fallback: this machine is for local testing/iteration; heavier training runs stay on rented cloud compute (RunPod) as needed.
• CPU and RAM sequencing: GPU was fixed first as the binding constraint, then CPU (Ryzen 9 9900X) and RAM (32GB/6000MHz) upgraded as reasonable-cost additions within remaining budget.
• Storage split, deliberate: the 1TB M.2 NVMe is for everyday/active-project use only. An old HDD will be added for local bulk storage, with archival data ultimately living on the home server.
• OS: Linux, matching the native OS already used for the robotics stack.

Open items:

• ✅ Linux install + environment setup — Ubuntu 26.04 LTS, Nvidia driver 595.84, CUDA 13.2 confirmed via nvidia-smi, Python 3.14.4, git, VS Code, GitHub SSH access all working. Full write-up: New ML Desktop — Setup Summary.
• ✅ Add old HDD for local storage; confirm home-server archival path — 2× 1TB Seagate HDDs added (repurposed from spare stock), plus the home server itself got a proper dedicated role and storage plan: Home Server — Setup Summary & Storage Plan.
• ☐ Validate actual training throughput on this GPU/RAM combo before resuming the SmolVLA plan in earnest

On the 10" touchscreen (currently idle on the Pi 3)

Worth keeping in reserve — good candidate uses once further along:

• A live dashboard Pi (showing Vision Kit's camera feed + detections during arm demos)
• A simple GUI for triggering pre-recorded voice/arm actions without needing to actually speak, useful for debugging
• Monitoring output for whichever Pi ends up as the "brain" coordinating the other two

Not needed for Phases 1–3, but there's no reason to buy a second display when this one's sitting idle — repurpose it once Phase 4 or 5 is reached.