Orbbec Gemini 2 + reBot Open-Source Robotic Arm: A Low-Cost Embodied AI Development Kit

This blog introduces a complete, low-cost, and fully hackable embodied AI learning kit built on Gemini 2 + reBot, and walks you through a hands-on robotic arm visual grasping demo.

1. Why Gemini 2 is Better for Robotic Arm?

We compare the Orbbec Gemini 2 with cameras like the Intel D435i and show how they perform across different environments. The Orbbec Gemini 2 is cheaper and its accuracy is comparable to the Intel Realsense D435i. Its SDK is perfectly compatible with OpenCV.

1.1 Camera Parameter Comparison Table

Camera	Technology	Depth Res	RGB Res	Range	FOV (D)	Weight	Interface	Price
Orbbec Gemini 2	Active Stereo IR	1280×800@30fps	1920×1080@30fps	0.15–10m	101°	98g	USB 3.0 Type-C	$240
Orbbec Gemini 335LG	Stereo Vision (Active IR)	1280×800@30fps	1280×800@60fps	0.17–20m+	~110°	164g	GMSL2 + USB-C	$449
Orbbec Gemini 336	Stereo Vision (IR-pass)	1280×800@30fps	1920×1080@30fps	0.10–20m+	~110°	99g	USB 3 Type-C	$310
SLAMTEC Aurora S	AI VSLAM + Stereo Fisheye	120° HFOV@15fps	180° Fisheye RGB	N/A (VSLAM)	180°	~200g	USB / Ethernet	$800
Intel Realsense D435i	Active IR Stereo	1280×720@30fps	1920×1080@30fps	0.28–10m	95°	72g	USB 3 Type-C	$419
Intel Realsense D405	Stereo (Short-range)	1280×720@30fps	1280×720	0.07–0.5m	~104°	~40g	USB 2.0/3.0	$419

1.2 Orbbec Gemini 2: Built for Dynamic Robotic Workflows

Dimension	Specification	Practical Value
Depth Technology	Active Stereo IR + IR speckle projector	High depth matching accuracy on low-texture surfaces, stable point cloud output
Shutter Type	Global Shutter	No rolling-shutter distortion during fast arm motion or object movement, excellent spatial consistency
Built-in Sensor	6-axis IMU	Stable point cloud output under motion, ideal for dynamic workflows
Depth FOV	101° diagonal (91° H / 66° V)	Covers ~0.5m × 0.35m at 0.3m working distance, perfect for desktop grasping
Range	0.15m – 10m, ideal 0.2 – 5m	Balances close-range precision with mid-range scene perception
Interface	USB 3.0 Type-C single cable	Data and power over one cable, plug-and-play, clean wiring
Power	2.0W avg (6.5W peak)	Powered directly from arm controller or embedded host, no extra PSU needed
Advanced	Multi-camera hardware sync (8-pin) + D2C depth-color pixel alignment	Enables multi-view fusion, hand-eye calibration, and other advanced applications
Summary	Balances real-world accuracy and development cost	No need to stack industrial specs for lab R&D — scenario-fit accuracy matters more than spec stacking

1.3 Real-World Camera Performance Comparison

Standard indoor lighting: Both Gemini 2 and D435i achieve better than ±2% depth accuracy at 2m, with similar point cloud density. Gemini 2’s 1280×800 resolution is slightly higher than D435i’s 1280×720, yielding sharper object edge details. The D405 is limited to a 0.5m max range, suitable only for close-range desktop inspection.

Low-light and low-texture: In dark environments (<50 lux) or on low-texture surfaces like white desks, passive stereo systems produce depth holes. Gemini 2’s active IR speckle projector significantly improves low-texture matching, and its IR global shutter maintains a stable 30fps even in the dark. The D435i also has active IR but with lower projector power, so its depth validity rate in very dark conditions is slightly below Gemini 2. The Gemini 336 uses an IR-pass filter for stronger ambient-light rejection under active IR projection, making it suitable for semi-outdoor sunlight scenarios.

Bright light / semi-outdoor: Under direct sunlight (>10000 lux), the D435i’s IR projector is overwhelmed and its valid depth area shrinks significantly. Gemini 2 is officially rated for indoor/semi-outdoor use and maintains ~70% depth validity in shaded semi-outdoor conditions.

Close-range precision: The D405 delivers sub-millimeter accuracy in the 7–50cm range, making it the best choice for small-object sorting. Gemini 2’s 0.15m minimum range, combined with its 1280×800 high resolution, already meets millimeter-level positioning needs in the typical 0.2–0.5m arm grasping distance, while offering much wider range coverage.

2. Vision Grasping Project: Orbbec Gemini 2 + reBot Arm

2.1 Project Features

1. Direct grasp pose estimation from YOLO + OBB: The pipeline uses detection boxes or OBB minimum-area rectangles directly and takes the short axis as the gripper opening direction, avoiding complex 3D point-cloud processing.
2. GraspNet-Baseline 6D grasp pose estimation (optional): The project also supports GraspNet-Baseline (graspnet/graspnet-baseline) for 6D grasp pose estimation from RGB-D point clouds, with YOLO bounding boxes used to select target candidates for more complex grasping experiments.
3. Robotic arm and gripper driver integration: The main grasping script is based on the arm and end-pose controller from reBotArm_control_py, with a lightweight grasp helper for gripper opening, force-controlled grasping, and TCP pose reading.
4. Open source and extensible: reBot is a 6+1 DoF open-source arm with 754mm max reach, 2.5kg rated / 5kg max payload, and ±0.1mm repeatability. Hardware, software, CAD drawings, SDK, BOM, and core algorithms are fully open source with no closed barriers — users can customize control algorithms and behavior to their needs.

2.2 Getting Started

• Detailed wiki tutorial: https://wiki.seeedstudio.com/rebot_arm_b601_rs_grasping_demo

2.3 reBot Core Specifications:

Specification	Value
DoF	6+1
Max Reach	754mm
Rated / Max Payload	2.5kg / 5kg
Repeatability	±0.1mm
Weight	6.5kg
Operating Temp	-20℃ ~ 50℃
Power	48V 15A

reBot Ecosystem Compatibility:

Ecosystem	Support	Practical Value
ROS1 / Moveit1	✅ Full	Compatible with legacy robotics frameworks, supports existing project iteration
ROS2 / Moveit2	✅ Full	Next-gen robotics standard, supports long-term project development
Python	✅ Full	Low-barrier hacking, fast algorithm debugging and scripting
LeRobot / Isaac Sim	✅ Full	One-click sim-to-real model transfer, solves the core sim-to-real gap
Pinocchio	✅ Full	Deep research and validation of kinematics and dynamics algorithms
Full Open Source	✅ Complete	Entire pipeline hackable, suitable for academic research and commercial prototyping

About Author