Voice-Controlled Grasping Arm
Say "Hey Jarvis, grab the water bottle" — the arm looks at the table through its wrist-mounted RGB-D camera, finds the bottle, plans a grasp and picks it up, then tells you what it did. Everything runs on the Jetson: wake word, speech recognition, the LLM that parses your intent, object detection, grasp planning and speech synthesis. No cloud, no online API.
What it can grasp
| Object | Strategy | Status |
|---|---|---|
| Cardboard boxes | side-face geometric grasp (multi-frame median, force 1.0 N·m) | verified |
| Cups (opaque, short) | side-face grasp, adaptive force | verified |
| Standing bottles (opaque) | cylinder route: level side approach, mid-body grip, fixed 0.8 N·m | verified |
| Bananas | elongated route (grip across the long axis) | verified |
| Oranges | round route: level approach, equator-height grip | verified |
| Transparent bottles | — | not possible: stereo depth cannot see clear plastic + water |
Objects too wide for the 0.100 m jaw get a spoken decline ("The box is too big for me to grip"). Detection runs on the GPU through a prebuilt native TensorRT engine: scene capture takes 0.6–1.6 s and a full voice-to-carry grasp cycle about 11 s.
How it works
reSpeaker mic ─▶ wake word ─▶ streaming ASR ─▶ Qwen3.5-4B (TensorRT-Edge-LLM)
│ grasp_object("water bottle")
▼
Orbbec Gemini2 ─▶ YOLOE open-vocab detector (10 classes)
│ instance mask
▼
depth cloud ─▶ PCA shape descriptor (elongation /
planarity / spine-bend) ─▶ route: box faces |
cylinder | elongated | round ─▶ 6-DoF grasp pose
│
▼
reBot B601-DM: approach ─ grip (force-controlled)
─ lift ─ carry home ─▶ TTS confirmation
Three containers, one compose file:
- rebot-arm — the agent: wake word, camera, detection, grasp pipeline, arm control, dashboard (
:8776) and observation API (:8775) - seeed-voice — streaming Qwen3 ASR + MOSS-TTS-Nano speech synthesis (
:8621) - edge-llm — Qwen3.5-4B-AWQ on TensorRT-Edge-LLM with MTP speculative decoding, ~45 tok/s (
:8000)
The detector takes its class vocabulary as a runtime input rather than baking it into the weights, so extending the recognizable object list is a config edit — no model re-export, no retraining.
The one manual step
Hand-eye calibration (one-time, ~30 min): grasping needs millimeter-accurate camera-to-arm geometry, which is physically unique to each unit. The guide walks through collecting ~16 poses of a printed ArUco board and solving for the transform. Until it's done, voice chat, detection and the dashboard all work — only grasping waits.
Requirements
- reComputer J4012 / Jetson Orin NX 16 GB (JetPack 6) — the compose file mounts host CUDA/TensorRT
- reBot B601-DM arm (USB serial) + Orbbec Gemini 2 on the wrist (USB 3.0)
- reSpeaker USB mic + any speaker
- ~8.5 GB of model downloads on first boot (LLM engine, speech engines, detector), plus ~1.4 GB of container images