A ten-station robot learning facility at Sonoma State University. Each station pairs a manipulator with a GPU workstation, so a student can record demonstrations, train a policy, and run it on the arm.
The purpose of the lab is to close the loop between learning and hardware. Policy work that stops at simulation does not tell you whether a method survives contact, latency, and a gripper that does not quite close where the model thought it would. Each station here is configured so that the same machine collects the data, trains the model, and drives the arm.
Six-axis research manipulator built for machine learning work, driven over Ethernet from a Python API rather than through a motion controller.
| Degrees of freedom | 6 |
| Reach | 700 mm, 1400 mm span |
| Payload | 1.5 kg |
| Repeatability | 1 mm |
| Joint torque | 27 N·m base three, 7 N·m wrist three |
| Gripper | 40 mm travel, 100 N grip force |
| Interface | Ethernet, 24 V |
Five-axis arm on DYNAMIXEL X-series smart servos, with the full Interbotix ROS 2 stack and a URDF that matches the hardware.
| Degrees of freedom | 5, with wrist rotate |
| Reach | 450 mm, 900 mm span |
| Working payload | 100 g |
| Repeatability | 2.5 mm |
| Servos | 6 total: XM430-W350 and XL430-W250 |
| Interface | USB, ROS 2 control |
| GPU | NVIDIA RTX 5080, 16 GB, Blackwell |
| CPU | AMD Ryzen 9 9900X, 12 cores |
| Memory | 32 GB DDR5-6000 CL30 |
| Storage | 1 TB PCIe Gen 5 NVMe |
| OS | Ubuntu 24.04 LTS |
| Networking | 5 GbE, plus a dedicated arm link |
| Intel RealSense D435 | Stereo depth and RGB |
| Luxonis OAK-D W | Stereo depth with on-camera inference |
| Fiducials | AprilTag detection for frame registration |
| Point clouds | Open3D processing pipeline |
| Capability | Stack |
|---|---|
| Robot middleware | ROS 2 Jazzy Jalisco, RViz, the Interbotix X-Series ROS 2 stack, and URDF models that match the hardware |
| Arm control | Trossen Arm Python SDK for the WidowX AI, interbotix_xs drivers for the RX150, joint and Cartesian control on both |
| Deep learning | PyTorch with CUDA on Blackwell, scikit-learn, and the full NumPy and SciPy stack |
| Kinematics | modern_robotics for screw-theory and product-of-exponentials work, alongside the analytic solvers in the vendor stacks |
| Simulation | MuJoCo for contact-rich policy training, Gazebo Harmonic for ROS 2 integration |
| Perception | OpenCV, Open3D, AprilTag detection, and the native RealSense and DepthAI Python bindings |
| Photoreal simulation | NVIDIA Isaac Sim and Isaac Lab, for large-scale parallel environments and sim-to-real transfer on the same GPUs that run the policies |
The lab is the teaching space for robotics at Sonoma State, and the learning here is experiential. Students develop the theoretical models, test them on state-of-the-art hardware, and build the practical experience a career in the robotics industry requires.
Projects are run by undergraduates and graduate students.
The lab is open to collaboration with companies and institutions in the region, and to faculty, students, and departments across Sonoma State.
Salazar 2008, Sonoma State University, Rohnert Park. Tours and collaboration enquiries are welcome.