QBot Platform
The mobile platform combines onboard GPU computing with an RGB-D camera, LiDAR, and inertial sensing. These resources support projects in image-based perception and learned decision-making alongside navigation and feedback control.
AVRC Laboratory · Unit 04
Artificial Intelligence
Machine learning, computer vision, and cooperative robotics connected to shared QBot Platform, QDrone 2, and QArm resources.
Artificial intelligence connected to physical systems
The AI unit brings machine learning, deep learning, and computer vision into AVRC’s shared robotics and autonomous-systems environment at Howard University, Department of Electrical Engineering and Computer Science, College of Engineering and Architecture.
Projects can connect sensor data to trained models, then evaluate how their predictions affect a measured robotic task. Python provides a development workflow; TensorFlow can support neural-network models, and ROS provides communication and integration tools. Each project must match the platform’s installed operating system, drivers, and software versions.
QDrone 2
The aerial platform combines onboard GPU computing with multiple cameras and supported software interfaces for programming and robotic-system integration. Candidate investigations include onboard visual inference and perception-assisted coordination.
QArm
The four-degree-of-freedom manipulator includes an RGB-D camera for color and depth observations. A connected computer can combine object recognition with calibrated geometry and motion planning for supervised manipulation experiments.
These are manufacturer platform specifications. Confirm the installed configuration before selecting a project. TensorFlow builds must match the JetPack release on each Jetson; software environments are not interchangeable between vehicles. NVIDIA TensorFlow compatibility guidance ↗
Possible AI projects and research directions
The following are proposed directions for faculty-supervised research, VIP teams, and senior design. Each begins with a baseline and a measurable question; they are not claims of completed AVRC demonstrations.
Perception for mobile robot navigation
How reliably can a robot recognize obstacles or target objects under changing indoor conditions?
A team could collect labeled RGB-D observations, evaluate a pretrained detector, and compare it with a simple image-processing baseline. Depth and LiDAR measurements can supply geometric information for a separate navigation controller. Test new lighting conditions and object arrangements using held-out recordings before supervised, low-speed trials.
Evaluation: Report detection precision and recall, distance-estimation error, inference latency, and navigation success under stated test conditions.
Air–ground leader–follower coordination
Can an aerial robot maintain a controlled relative position as a ground robot follows a prescribed path?
Begin with a simulated QBot leader and aerial follower, then establish a position-based tracking baseline using a shared coordinate frame. A possible AI extension is an image-based detector or learned estimator that helps locate the ground robot when visual observations are available. Compare that extension with the baseline while retaining bounded references and a verified flight controller; test communication delay and observation loss in simulation first.
Evaluation: Measure relative-position error, separation, end-to-end delay, and recovery after interrupted observations. Physical trials depend on a suitable tracking volume and an approved flight plan.
Cooperative air–ground inspection
When do complementary aerial and ground observations improve a small indoor inspection task?
A drone could identify a marked object or region from an elevated viewpoint and send its estimated location and confidence to a ground robot for closer inspection. The project would connect image recognition with time-stamped messages, coordinate transformations, and task assignment. Begin with recorded data and a stationary camera, then compare a coordinated mission with a single-robot baseline in a bounded laboratory scenario.
Evaluation: Evaluate localization error, false detections, task completion time, and the effect of stale or inconsistent observations.
Multi-agent coordination and swarm methods
How do local communication rules and learned decisions affect a group’s coordination?
Start with a small simulated network using established consensus, formation, or coverage rules. A learning extension could adapt task allocation or communication choices, with performance compared against those deterministic rules. Multi-agent reinforcement learning is a possible advanced study, but coordination itself does not require AI. Increase the number of simulated agents before considering a small physical demonstration with the available vehicles.
Evaluation: Record formation or coverage error, minimum separation, message load, and performance under delay or packet loss. A simulation study does not establish that a physical swarm has been validated.
Vision-guided sorting and manipulation
Can visual recognition make a calibrated pick-and-place task more adaptable?
Use the QArm camera to identify a small set of tabletop objects and estimate their positions from depth measurements. Calibrate the camera-to-robot transformation, then combine the visual estimate with inverse kinematics and an approved motion sequence. Compare a learned object classifier or detector with a color- or marker-based baseline, using reachable objects and grasp orientations compatible with the four-degree-of-freedom arm.
Evaluation: Measure classification accuracy, object-position error, grasp success, and cycle time across held-out object placements. Model inference can run on a connected computer; the arm is not described as having an onboard Jetson.
Learning policies and testing transfer to hardware
How much does a policy learned in simulation change when sensing and dynamics differ?
A team could investigate imitation learning or reinforcement learning for a bounded navigation task, using a simulator with a suitable training interface. Compare the policy with a conventional navigation baseline and vary sensor noise, latency, and dynamics during training and evaluation. Freeze the trained policy for evaluation; any hardware transfer would follow offline checks and supervised trials with independent stop and motion limits.
Evaluation: Report success rate, tracking error, constraint violations, inference latency, and the gap between simulated and physical results. QLabs can support preparation and comparison; a compatible reinforcement-learning interface must be established for the selected simulator.
Connections to published research
Fadel Lashhab and collaborators have published work on data-driven sensing, UAV deployment, and dynamic consensus networks. These studies provide context for the proposed investigations above; each new hardware experiment still requires its own validation.
Data-Driven Design for IRS-Assisted Massive Multi-Modal Sensing in Industry 5.0
M. Haider, I. Ahmed, F. Lashhab, T. O’Shea, D. B. Rawat, M. A. Matin
2024 IEEE International Conference on Communications Workshops (ICC Workshops)
Optimization of UAV Deployment for Enhanced Communication Coverage
T. N. Islam, F. Lashhab, I. Ahmed
American Journal of Science Education Research, AJSER-151, 2024
A practical software and evaluation workflow
- Establish the baseline. Use Python and the supported platform interfaces to collect time-stamped sensor data and reproduce a simple method before introducing a learned model.
- Train and test deliberately. Separate training, validation, and test data. A suitable workstation can train a model; the embedded Jetson can evaluate inference performance within its memory and timing limits.
- Integrate perception with control. Use the supported ROS or ROS 2 workflow where appropriate. Verify message timing and coordinate frames, and retain a known controller with bounded references.
- Evaluate transfer. Compare recorded-data, simulation, and supervised hardware results. Document failures and the limits of the test conditions, including sensor noise and communication delays.
Technical webinars on AI and cooperative robotics
Official Quanser recordings provide technical context and examples from outside AVRC. The coordination webinar presents a control approach; it is a useful baseline for a later learning extension. Select a video to load the player.
3 of 3 videos · Official Quanser channel
AI & cooperative robotics
Applied AI with Quanser Autonomous Vehicles
Quanser’s Python Intelligence Toolset and image-based neural-network examples for GPU-enabled autonomous platforms.
Try a learning prompt
Separate training from inference, identify the sensor inputs, and propose a latency and accuracy test.
AI & cooperative robotics
Closing the Sim-to-Real Gap: Inside Quanser’s Reinforcement Learning Research
A September 2026 account of an ongoing QBot navigation study, including an Isaac Lab training workflow and unresolved transfer challenges.
Try a learning prompt
Identify one source of simulation-to-hardware mismatch and design a test that could isolate it.
AI & cooperative robotics
Cooperative Decentralized Control of Heterogeneous Multi-Vehicle Systems
An August 2026 webinar on model-based formation control with aerial and ground vehicles. This provides a coordination baseline; it is not presented as a machine-learning method.
Try a learning prompt
Which relative measurements are needed? Explain where a learned perception method might assist the controller.
Bring your course or project to the lab.
Discuss the learning goals, platform, team size, and supervision needed for a practical lab experience.