Benjamin Aziel

I'm a roboticist (primarily) interested in control and planning for autonomous robotic systems. I'm committed to building systems that are both capable and dependable in constrained, uncertain environments.

About

I recently completed an M.S. in Mechanical Engineering and Applied Mechanics at the University of Pennsylvania, with a concentration in Mechatronic and Robotic Systems. I graduated summa cum laude with a B.E. in Mechanical Engineering from The Cooper Union , where I also minored in Computer Science.

My research interests center on developing control and motion planning algorithms with rigorous guarantees for safety, stability, and reliability in uncertain and evolving environments. I'm particularly interested in unifying tools from dynamical systems and optimal control with data-driven modeling, with the goal of enabling robots to adapt online without sacrificing analytical structure or formal guarantees.

At Penn, I conducted research in the Sung Robotics Lab within the GRASP Lab, where I worked on motion generation and path planning for the NASA-funded TRUSSES lunar robotics project under the guidance of Dr. Cynthia Sung and Dr. Dan Koditschek. My work explored structure-preserving formulations for risk-aware planning, including energy-based models that explicitly incorporate velocity sensitivity.

Select Projects

Whole-Body Control for Quadruped Locomotion on the Unitree Go2

Benjamin Aziel

I designed an instantaneous QP-based whole-body controller for trotting locomotion on the Unitree Go2, implemented in MuJoCo. At each control step, a single QP resolves contact forces and joint torques simultaneously, satisfying rigid-body dynamics, friction cone constraints, and actuator limits while tracking Cartesian tasks for base height, orientation, and velocity. Swing foot trajectories are generated online with a Raibert-style heuristic, and early touchdown detection handles contact timing mismatches during the trot gait.

Implementation of Operational Space Control Barrier Functions on Franka Manipulator

Benjamin Aziel, Mateusz Jaszczuk

We implemented a safety-critical control framework integrating control barrier functions into an operational space controller for a 7-DoF Franka Emika Panda, enforcing joint limit, obstacle avoidance, and self-collision constraints via high-order CBFs in a QP formulation. The controller was implemented using OSQP and simulated with PyBullet; it sustained ~1.3 kHz in headless simulation, well surpassing the 1 kHz real-time requirement of the arm.

Learning Residual Vehicle Dynamics for MPPI

Benjamin Aziel, Brian Fok, Anuriha Kodali

We augmented a kinematic bicycle model with a learned neural residual following the Knowledge-based Neural ODE (KNODE) framework and integrated it into an MPPI controller on an F1Tenth racecar. Despite a 92% reduction in velocity error and 31% reduction in positional drift in open-loop rollout, closed-loop tracking did not improve (one-step MSE training has no guarantee of preserving the relative cost ranking MPPI induces over hundreds of sampled trajectories across a planning horizon).

Passive SE(3) Energy Tank Control for Contact-Rich Manipulation

Benjamin Aziel (with guidance from Davin Tjandra)

I designed a passivity-enforcing controller for 6-DoF contact-rich manipulation by extending energy tanks to full SE(3), incorporating rotational dynamics in quaternion space alongside the standard translational tank. The dual-tank architecture gates non-passive force and torque commands by available stored energy, guaranteeing the closed-loop system never generates net energy. Validated on a KUKA LBR iiwa 7 with an OptoForce wrist sensor on a J-groove insertion and path-following task.

Vision-Guided Pick-and-Place with 7-DoF Franka Manipulator

Benjamin Aziel, Mateusz Jaszczuk, Andrik Puentes, Solomon Gonzalez

We developed a robotic system for autonomous block detection, grasping, and stacking in both static and dynamic environments, built on the Franka Research 3. The pipeline combined AprilTag-based perception from a wrist-mounted camera, numerical inverse kinematics with null-space projection, and a dynamic interception strategy using angular velocity estimation for moving targets.