# HumanHUD desktop robot arm: procurement and integration plan

Research checked September 21, 2026. Prices are USD, before shipping, tax and import charges. No purchase has been made. This is a prototype plan, not a claim that HumanHUD currently controls this hardware.

**Recommended first platform: one SO-ARM101 Pro leader–follower pair, clamped to a dedicated desk.** It gives MathaLabs a repairable, 3D-printable manipulation platform and an existing demonstration-recording path. Start with enterprise workstation tasks: sorting small parts into trays, presenting a lightweight component, and moving small objects between marked locations. Build the robot backpack after desk demonstrations and operator controls are validated.

## What to buy

The fastest starting point is [Seeed SO-ARM101 Pro Assembled Kit, SKU 100046482](https://www.seeedstudio.com/SO-101-Assembled-Kit-Pro-p-6691.html), listed **$299.00, in stock**. Its detailed parts list confirms both assembled arms, flexible gripper parts, a 32 × 32 mm USB camera, two driver boards, separate 5 V and 12 V supplies, two USB-C cables, power pigtails, four clamps, screwdriver and desk mat. A computer is additional. This preserves the option to print replacement or custom frames later.

For a kit to assemble or print yourself:

| Item | Quantity | Current unit price | Subtotal | Included / purpose |
|---|---:|---:|---:|---|
| [Seeed SO-ARM101 Pro Servo Motor Kit, 114993667](https://www.seeedstudio.com/SO-101-Low-Cost-AI-Arm-Kit-Pro-p-6427.html) | 1 | $249.90 | $249.90 | In stock; electronics for leader + follower, boards, power and cables; no arm frames or camera |
| [SO-101 printed parts, 114993668](https://www.seeedstudio.com/SO-101-3D-printed-Enclosure-p-6428.html) | 1 | $30.99 | $30.99 | In stock; both arms, clamps and flexible gripper parts; omit if self-printing |
| [X10 1080p USB camera, 114090066](https://www.seeedstudio.com/X10-USB-wired-camera-p-6506.html) | 2 | $12.00 | $24.00 | In stock; one fixed desk view and one secondary view; determine mounting before ordering |
| **Assembly route hardware subtotal** | | | **$304.89** | Existing computer assumed |

The assembled route plus one additional X10 camera totals **$311.00**. Do not buy the electronics kit and assembled kit together; these are alternative routes. The separate parts page is $30.99 even though some accessory widgets show $29.90; use the actual item price and verify the final cart. The site's static comparison graphic contains older prices, so it is not used for these totals.

Reserve an additional **$75–150 planning allowance**, not a vendor quote, for a suitable powered USB hub, desk camera mount, cable management, basic tools and spare fasteners. That puts the assembled bench at approximately **$386–461 before shipping/tax**, with an existing Mac or Ubuntu computer. A $500–1,000 development budget leaves room for spare parts and iteration; a new training computer or printer is outside that total.

## Alternatives and availability

| Option | Verified listing | Assessment |
|---|---|---|
| SO-ARM101 Pro, Seeed | $299 assembled pair; in stock | Recommended; straightforward stock hardware and explicit pair contents |
| [SO-ARM101, PartaBot](https://partabot.com/products/so-arm101) | Electronics $329; full unassembled $399; assembled $479 | US shipping origin. Official [product variant feed](https://partabot.com/products/so-arm101.js) reports all three available; page also contains a generic “Sold out” label, so confirm the selected variant at checkout. Full kits include printed PLA+ frames. |
| [SO-ARM101 No-3DP CAM Kit (SE), Waveshare SKU 34807](https://www.waveshare.com/so-arm100-3dp-parts-kit.htm?sku=34807) | $229.99; structured listing says in stock | Lower-cost candidate with camera accessories, without frames. SE leader uses encoder-only virtual servos; follow [vendor-specific setup](https://www.waveshare.com/wiki/SO-ARM100/101), not assumed stock leader motor settings. |
| [SO-ARM100, Seeed](https://www.seeedstudio.com/SO-ARM100-Low-Cost-AI-Arm-Kit.html) | $200 electronics; discontinued/out of stock | Do not start a new build here. [Upstream](https://github.com/TheRobotStudio/SO-ARM100) marks SO100 documentation deprecated; SO101 improves wiring and leader assembly. |
| [ROBOTIS OpenMANIPULATOR-X](https://robotis.us/products/openmanipulator-x-rm-x52-tnm) | $1,629.09 arm; page says sold out and 2–3 week lead time | Above proposed budget before separate 12 V power, controller/interface and base. Published 4 arm DOF + gripper, 500 g payload, 380 mm length. Useful later if ROS/DYNAMIXEL becomes a requirement. |

## Components, printing and desk installation

The SO101 Pro motor set has six 12 V follower servos and six 7.4 V-rated leader servos: one 1:345, two 1:191 and three 1:147 gear ratios. The leader operates from the supplied **5 V** source; follower from **12 V**. Keep these supplies labeled and separate throughout setup. Two USB servo-control boards connect the arms to the host computer. Follow the [Seeed motor mapping and power instructions](https://wiki.seeedstudio.com/lerobot_so100m_new/) for the exact kit revision.

Print structural links, bases, gripper parts and the appropriate controller mount; buy motors, boards, cables, power and fasteners. The [official design repository](https://github.com/TheRobotStudio/SO-ARM100) includes STEP, STL and simulation descriptions. The [printing guide](https://github.com/TheRobotStudio/SO-ARM100/blob/main/3DPRINT.md) identifies shared parts needed twice for a pair, leader/follower-specific parts, and board-specific mounting plates. Use individual printable files, not the complete assembly STL. Its print-service guidance uses PLA+ and 20% infill. An existing FDM printer or a quoted printing service is sufficient; buying a printer is not required.

For assembly, have precision screwdrivers/hex keys matching the fasteners, a ruler or caliper, labels, and a stable desk. Clamp both bases and verify desk thickness against the physical clamps. Leave the moving follower in a separate marked workspace with cables outside its sweep. Do not drill a production mounting plate from a guessed pattern: take dimensions from the selected STEP model and confirm them on the received kit.

Seeed's product specification advertises **500 g payload and 5 + 1 DOF**. This is a vendor claim with no pose/duty-cycle qualification found; it is not a validated HumanHUD operating envelope. Start with lightweight objects, measure loaded reach and repeatability, and publish a task-specific limit only after bench testing. No verified reach or mounting-hole dimensions are asserted here. Servo torque in kg·cm must not be substituted for whole-arm payload.

The [vendor operating guide](https://wiki.seeedstudio.com/lerobot_so100m_new/) calls for a firmly fixed arm, clearance around its motion area, and stopping/powering off on abnormal movement or communication problems. Its stated 1 m clearance makes a dedicated workstation preferable to an occupied everyday desk. An accessible power disconnect, low-speed commissioning and a local operator are requirements of this proposed prototype. “Stop” in software is not a certified emergency-stop circuit.

## Host software and licensing

Use the official [LeRobot SO101 setup](https://huggingface.co/docs/lerobot/so101) for motor IDs, calibration and ports. It documents macOS USB serial devices; [installation guidance](https://huggingface.co/docs/lerobot/installation) covers macOS/MPS and Linux/CUDA. A Mac can perform control and recording. Ubuntu with NVIDIA GPU is a practical training target; the [real-world tutorial](https://huggingface.co/docs/lerobot/main/getting_started_real_world_robot) also provides MPS training examples. Exact policy support and speed depend on hardware and dependencies. No Jetson is needed for the first USB-connected bench.

Pin a tested LeRobot release/commit and dependencies before commissioning; current main installation instructions require Python 3.12+ and PyTorch 2.10+. Install the robot workflow and Feetech extras, then calibrate the actual hardware. Keep separate calibration files for leader and follower and record their hashes with demonstrations.

The [SO-ARM design repository](https://github.com/TheRobotStudio/SO-ARM100/blob/main/LICENSE) is Apache-2.0. [LeRobot's license](https://github.com/huggingface/lerobot/blob/main/LICENSE) is Apache-2.0 with notices for included third-party code. Preserve applicable notices when redistributing modifications. Do not assume a purchased servo's firmware or every pretrained model/dataset shares that license.

## Proposed HumanHUD integration milestones

1. **Commission the bench.** Photograph the received BOM, record motor variants and serial ports, clamp and calibrate both arms, and verify manual leader–follower control. Pass condition: repeatable limited-motion tests and reliable local stop/disconnect behavior.
2. **Create a local robot adapter.** Put hardware access behind an authenticated local service exposing status, calibration identity, bounded motion commands, gripper control and stop. Permit one controller lease. Reject expired commands, invalid units, uncalibrated operation and targets beyond configured joint/speed bounds. Keep motor control on the workstation; glasses send intent and display state.
3. **Connect HumanHUD interaction.** Start with deliberate “select object → preview action → hold to execute” interaction and visual robot status. Treat voice/vision detections as suggestions until explicitly armed. Loss of connection or operator release cancels the command stream. Use a fixed camera and known tray coordinates before adding gaze-based target estimation.
4. **Capture linked episodes.** Record robot observations/actions, fixed and wrist camera frames, HumanHUD event timestamps, task descriptions and operator interventions. Add egocentric streams only where the glasses API and participant permissions allow capture. Synchronize clocks and retain measured timing error; visual activity labels alone are not robot action labels.
5. **Curate one enterprise task.** Label successful/failed small-part transfers; review boundaries, object identities and recovery behavior. Split training/evaluation by session/operator rather than neighboring frames. Export a versioned dataset manifest, provenance, consent scope and calibration metadata alongside LeRobot data.
6. **Evaluate assistance before autonomy.** Report task completion, intervention rate, latency, drops and failure cases on held-out trials. Train a small imitation-learning baseline only after a clean demonstration set exists. SAM masks/tracks may assist annotation; they do not supply physically valid actions. Simulation-derived records must remain explicitly distinct from real demonstrations.

The first customer package should therefore be a supervised workstation demo plus reproducible task data and integration SDK. Backpack, mobile cleaning and yard concepts remain later hardware research until mounting loads, power, weather exposure and control requirements are separately validated.