Wearable data collection kit for training a humanoid robot #17

BotshareAI · 2026-04-04T23:05:40Z

BotshareAI
Apr 4, 2026
Maintainer

Technical description

Project overview

This project develops a wearable data collection kit for training a humanoid robot through multimodal human demonstration capture. The system is built around a backpack-mounted compute and power unit, a helmet-mounted stereo perception module, and left/right wrist-mounted camera modules for hand-level observation. The backpack contains the main compute platform, storage, battery system, power distribution, monitoring electronics, and connector interface, while the wearable mounts are designed for quick installation, reliable cable routing, low operator fatigue, and field serviceability. The target architecture is modular, production-aware, and suitable for repeated daily use, with support for synchronized video, inertial sensing, GNSS, battery status, thermal monitoring, and safe battery swap during operation.

The external-frame backpack selected for the project is a good fit for this concept because it naturally supports load distribution, equipment mounting, airflow, and service access around the central electronics bay.

Assumptions

The ZED IMU gives motion at the helmet / head reference frame. It does not give the orientation or angular motion of the left and right forearms. If downstream pipeline only needs:

egocentric stereo vision
operator hand-view video
GNSS
timestamps

then wrist IMUs are optional.

If downstream pipeline also needs:

left/right wrist orientation
arm motion priors
better frame-to-frame motion compensation for hand cameras
easier alignment of human motion with robot arm trajectories

then adding one IMU per wrist camera set is a good idea.

Recommendation

For the first working prototype:

keep wrist IMUs as an optional but recommended module
design the mechanical wrist pod so an IMU can be added without redesign
include them if your dataset is meant for humanoid manipulation learning rather than only video logging

Recommended design philosophy

The goal is to minimize custom electronics, build the first prototype almost entirely from ready-made modules and avoid a custom carrier board.

Preliminary component and power table

Hardware components list

Subsystem	Component	Supply rail	Peak current	Typical power	Notes
Main compute	NVIDIA Jetson Orin Nano Developer Kit	12 V or native input	up to 3.3 A at 12 V	10 - 20 W typical, up to 40 W in Super mode	Supported power modes include 7 W, 15 W, 25 W, and up to 40 W in Super mode. Board
Main stereo camera	Stereolabs ZED X Mini GMSL	12 V	0.5 A	4 - 5 W	Built-in IMU included. Camera
Left wrist camera	Arducam 16MP Wide Angle USB Camera with Case	5 V	0.4 A budget	1.5 - 2 W	Use 0.5 A budget for safe power design. Option 1, Option 2
Right wrist camera	Arducam 16MP Wide Angle USB Camera with Case	5 V	0.4 A budget	1.5 - 2 W	Use 0.5 A budget for safe power design. Option 1, Option 2
Left and right wrist IMU					Optionally
4/5G internet module					Optionally
GNSS	SparkFun GPS Breakout - NEO-M9N, Chip Antenna (Qwiic)	3.0 - 3.3 V	up to 100 mA peak	about 0.08 - 0.11 W typ.	GPS Breakout with on-board chip antenna, GPS, GLONASS, Galileo, and BeiDou constellations witn ~1.5 meter accuracy
Storage	NVMe SSD 1 - 2 TB	3.3 V / M.2 rail	1.2 - 1.5 A	2 - 3 W typical, 4 - 5 W peak	Use low-power NVMe to improve power consumption. SSD
Jetson cooling	40 - 60 mm PWM fan / blower	5 V or 12 V	0.3 A	1 - 2 W	Depends on enclosure airflow and dust filter. Already included above.
USB hub (optionally)	Industrial powered USB 3.0 hub	5 V	depends on loads	0.5 - 1.5 W self + device load	Recommended if both wrist cameras connect to one robust hub in the backpack. Hub
GMSL deserializer / carrier	GMSL interface board	12 V / 5 V / 3.3 V	0.4 A	1 - 3 W	GMSL Stereolabs card mono.
DC-DC converter, 14.8V to 12V	Main buck regulator	14.8 V input	system-dependent	conversion loss only	Required if the battery bus is 4S Li-ion. SD-50A-12
DC-DC converter, 14.8 V / 12 V to 5 V	Secondary buck regulator	input bus	system-dependent	conversion loss only	Powers webcams, hub, sensors, some accessories. SD-25A-12
Power distribution board with fuses					PDB, eWiring_Book (attached)
Hot-swap / ideal diode stage	LTC4412 / hot-swap controller class	battery bus	N/A	small loss	Needed for safe battery swap without Jetson brownout. Option_1, Option_2
Hold-up buffer	supercapacitor or small buffer battery pack	system bus	burst only	burst only	Keeps the computer alive during battery swap transition. Option_1, Option_2
Main BMS	4S Li-ion BMS with balancing	battery bus	pack-dependent	protection only	Include overcurrent, undervoltage, balancing, temperature. UART->NVIDIA; DIO -> buzzer; BT -> Daly App Option_1
Active buzzer (battery and system status)	→ GPIO Jetson				Option_1

Practical interpretation

1. ZED IMU

The ZED X Mini already includes an IMU, so the extra IMUs could only be added to the left and right wrist camera sets.

2. Wrist camera set recommendation

Each arm set should include:

1 × Arducam 16MP Wide Angle USB Camera with Case
local strain relief
quick-disconnect cable
optional tiny local status LED

That gives you:

hand-view video

3. Other components

powered USB hub
GMSL interface board
NVMe SSD
4S battery pack
BMS
hot-swap controller
hold-up buffer
14.8 V to 12 V converter
12 V / 14.8 V to 5 V converter
battery current / voltage monitor
temperature sensors (on BMS, Jetson)
fuse / e-fuse
fan / airflow path
locking connectors and cable strain relief

Estimated total system load

Operating case	Estimated total power
Moderate load	28 - 35 W
Heavy load	35 - 45 W
Worst-case design budget	45 - 55 W

That aligns with the 12-hour target in the proposal: at 35 W average, about 420 Wh is needed before margin, and about 500 Wh is a safer design target once converter losses and aging are included.

Recommended concise system architecture

Battery pack, 4S Li-ion, 14.8 V nominal
→ BMS + fuse + temperature sensors
→ hot-swap / ideal diode stage
→ hold-up buffer
→ 12 V rail for Jetson and ZED
→ 5 V rail for wrist cameras, USB hub, indicators
→ 3.3 V rail for GNSS, IMUs

Architecture for the Hot-swap

Main battery
│
BMS
│
ideal diode module
│
+---- system bus ---- supercap ---- regulators
│
ideal diode module
│
BMS
│
Backup battery

Block view:

Main battery
→ fuse → BMS → VIN1

Backup battery
→ fuse → BMS → VIN2

VOUT
→ supercap → distribution → rails

Battery configuration

Main battery

4S7P
28 cells total
nominal voltage: ≈14.8 V
nominal capacity: ≈35 Ah
nominal energy: ≈518 Wh

Dimensions:
≈150 mm × 85 mm × 75 mm

Weight:
≈2.3–2.5 kg

Backup battery

4S1P
4 cells
nominal voltage: ≈14.8 V
nominal capacity: ≈5 Ah
nominal energy: ≈74 Wh

Dimensions:
≈75 mm × 45 mm × 75 mm

Weight:
≈350–450 g

Links

Magnetic Pogo Pin connector UART + Battery

https://de.aliexpress.com/item/1005009493489658.html

https://de.aliexpress.com/item/1005010467009688.html

https://www.amazon.de/-/en/Spring-Loaded-Pitch-Female-Contact/dp/B0BXDFYZFG/

https://www.amazon.de/-/en/Waterproof-Magnetic-Positions-Spring-Contact/dp/B0DK8ZGQBN/?th=1

BotshareAI · 2026-05-17T00:11:40Z

BotshareAI
May 17, 2026
Maintainer Author

OpenAMRobot

Open Embodied AI Ecosystem

OpenAMRobot is an open, modular, and affordable embodied AI robotics ecosystem focused on:

Mobile Manipulation
Teleoperation
Imitation Learning
Vision-Language-Action (VLA) Robotics
Human-Centered Robot Policy Training

The project combines:

ROS 2 based robotics
Dual-arm mobile manipulation
Wearable AI data collection
AI training & simulation pipelines
Open-source hardware & software infrastructure

Open Embodied AI Pipeline

Human Demonstration
→ Wearable AI Kit
→ Dataset Generation
→ Policy Training
→ Simulation
→ ROS 2 Execution
→ Real-World Deployment
→ Feedback & Improvement

Core Architecture

1. Wearable AI Data Collection Kit

Human demonstration and teleoperation system designed for embodied AI research.

Components

RGB-D Head Camera
IMU Tracking (Head / Body / Wrists)
Optional Wrist Cameras
Optional EOG-based Gaze Estimation
Backpack Edge Compute Unit

Goals

Human Demonstration Recording
Teleoperation
Dataset Generation
Shared Autonomy
Robot Policy Training

2. OpenAMRobot Platform

ROS 2 based mobile manipulation platform.

Robot Components

Dual 6-DoF Arms
Mobile Base
Depth Cameras
RGB Cameras
3D LiDAR
IMU Sensors

Capabilities

Navigation
Manipulation
Teleoperation
Human-Robot Collaboration
Autonomous Task Execution

AI & VLA Stack

AI Training

Imitation Learning
Behavior Cloning
Policy Learning
Diffusion / Transformer Policies

Vision-Language-Action (VLA)

Vision
→ Language
→ Action

Current Direction

High-Level Policy Inference
Long-Horizon Task Execution
Shared Autonomy
Human-Centered Robot Learning

Technologies

NVIDIA GR00T (optional)
NVIDIA Isaac Lab (optional)
NVIDIA Isaac Sim (optional)

Compute Architecture

Low-Level Control

STM32 / Teensy / MCU
Motors
Encoders
Safety I/O

Mid-Level Compute

Raspberry Pi 5
Navigation
SLAM
ROS 2 System Management

High-Level AI Compute

NVIDIA Jetson Orin / Orin NX
Perception
Manipulation
VLA / Policy Inference

Wearable Edge Compute

RK3588-based Modules (optional)
Sensor Fusion
Data Synchronization
Streaming & Compression

Software Stack

Robotics

ROS 2
MoveIt 2
Nav2
TF2
rosbag2

AI / ML

PyTorch
OpenCV
CUDA
HuggingFace Transformers

Simulation

Gazebo
NVIDIA Isaac Sim (optional)
NVIDIA Isaac Lab (optional)

Key Principles

Open Source
Modular
Affordable
DIY Friendly
Research Oriented
Human-Centered
Reduced Vendor Lock-In

Roadmap

Phase 1

Open Mobile Robot Platform

Navigation
SLAM
Mobile Manipulation

Phase 2

Teleoperation & Data Pipeline

Wearable AI Kit
Human Demonstration
Dataset Generation

Phase 3

Imitation Learning & Policies

Policy Training
Simulation Integration
VLA Experiments

Phase 4

Shared Autonomy & Embodied AI

Human-Robot Collaboration
Long-Horizon Tasks
Semi-Autonomous Manipulation

Phase 5

Open Ecosystem & Community

Open Datasets
Research Collaboration
Educational Infrastructure

Mission

Democratize embodied AI and robotics by building open, affordable, and scalable infrastructure for the next generation of intelligent robotic systems.

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

OpenAMRobot

Wearable data collection kit for training a humanoid robot #17

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 1 comment

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Select a reply

Uh oh!

OpenAMRobot

Wearable data collection kit for training a humanoid robot #17

Uh oh!

Uh oh!

BotshareAI Apr 4, 2026 Maintainer

Technical description

Project overview

Assumptions

Recommendation

Recommended design philosophy

Preliminary component and power table

Hardware components list

Practical interpretation

1. ZED IMU

2. Wrist camera set recommendation

3. Other components

Estimated total system load

Recommended concise system architecture

Architecture for the Hot-swap

Battery configuration

Main battery

Backup battery

Links

Magnetic Pogo Pin connector UART + Battery

Magnetic Pogo Pin connector Battery (Option)

Magnetic Pogo Pin Connectors

Backpack frame

Helmet

Wrist strap

Replies: 1 comment

Uh oh!

Uh oh!

BotshareAI May 17, 2026 Maintainer Author

OpenAMRobot

Open Embodied AI Ecosystem

Open Embodied AI Pipeline

Core Architecture

1. Wearable AI Data Collection Kit

Components

Goals

2. OpenAMRobot Platform

Robot Components

Capabilities

AI & VLA Stack

AI Training

Vision-Language-Action (VLA)

Current Direction

Technologies

Compute Architecture

Low-Level Control

Mid-Level Compute

High-Level AI Compute

Wearable Edge Compute

Software Stack

Robotics

AI / ML

Simulation

Key Principles

Roadmap

Phase 1

Phase 2

Phase 3

Phase 4

Phase 5

Mission

BotshareAI
Apr 4, 2026
Maintainer

BotshareAI
May 17, 2026
Maintainer Author