Michigan Robotics at IROS 2026 in Pittsburgh

Zachary Charlick

Nilay Roy Choudhury

Haoyu Ma

Xiaonan Huang

Dmitry Berenson

STEMbot is a miniature compliant climbing robot designed to navigate underneath plant canopies, where it can enable early detection of pests before infestations spread. The system integrates a fully geometric PIN-SLAM pipeline with a semantic OcTree to achieve robust localization and mapping while climbing plants, reliably traversing stems from 7 to 33 millimeters across four distinct plant specimens. The work to develop soft microrobots that climb plant stems to rescue crops from pests is funded by the USDA and led by Xiaonan Huang and Dmitry Berenson.

Luís Marques

Maani Ghaffari

Dmitry Berenson

When a robot’s model of its own dynamics is wrong, how can it still make guarantees about where it will end up? This work, published in IEEE Robotics and Automation Letters, proposes Conformal Lie-group Action Prediction Sets (CLAPS), a symmetry-aware algorithm based on conformal prediction that constructs, for a given action, a set guaranteed to contain the resulting system configuration at a user-defined probability. Experiments on a real MBot show that accounting for the structure of the configuration space leads to more volume-efficient prediction regions.

Yulun Zhuang

Yue Qin

Justin Lu

Zelin Shen

Yichen Wang

Sicheng He

Yanran Ding

Adding springs to a robot’s legs can make it faster and more efficient. But, designing the body and the controller together is notoriously hard, because contact dynamics and mechanism engagement are non-differentiable. SurGE, from Yanran Ding’s lab, computes surrogate gradients of the design objective through a differentiable pipeline and injects them into an evolutionary search. Starting from a hand-tuned initial design of a hopping robot with a parallel spring, SurGE reduced the design objective by 37.65% on physical hardware, with improvement trends identified in simulation transferring consistently to the real system.

Jeeho Ahn

Christoforos Mavrogiannis

How can a car-like robot tidy up a densely cluttered space when it can only push objects, not pick them up? This work, published in IEEE Robotics and Automation Letters and presented by robotics PhD student Jeeho Ahn, tackles multi-object rearrangement problems that combine kinematic, geometric, and physics constraints. Their framework, ReloPush-BOSS, exhibits consistently highest success rates and shortest pushing paths.

Yeheng Zong

Yizhou Chen

Alexander Bowler

Chia-Tung Yang

Ram Vasudevan

How can robots learn how deformable objects move when capturing 3D data on objects that bend, fold, and stretch typically requires labor-intensive annotation or expensive motion capture setups? TrackDeform3D, from the lab of Ram Vasudevan, professor of robotics, is an affordable and autonomous framework for collecting 3D datasets of deformable objects using only RGB-D cameras. The method identifies 3D keypoints and robustly tracks their trajectories, incorporating motion consistency constraints to produce temporally smooth and geometrically coherent data, and it outperforms state-of-the-art tracking methods in both geometric and tracking accuracy. Using the framework, the team built a high-quality, large-scale dataset covering six deformable objects and totaling 110 minutes of trajectory data.

Matteo Crotti

Elizabeth B. Wilson

Manuel Giuseppe Catalano

Elliott Rouse

Most powered prosthetic legs use a simple, non-adaptive foot, which is easy to build and control but can limit performance on uneven ground and around obstacles. This work from the lab of Elliott Rouse, accepted to the State of the Art in Robotic Leg Prostheses workshop on October 1, compares a powered prosthetic leg fitted with either a standard fiberglass foot or a mechanically adaptive one, testing both during walking with visible obstacles and on a treadmill with obstacle-like perturbations. The study suggests adaptability built into the foot itself can improve a prosthesis’s robustness without any changes to its higher-level control.

Andrew Choi

Dezhong Tong

Xiaonan Huang

Soft robots are notoriously slow to train in simulation because their contact-heavy dynamics bog down conventional simulators. This work from the lab of Xiaonan Huang, assistant professor of robotics, shows that fully implicit time-stepping in the DisMech soft-body simulator, paired with a new delta natural curvature control scheme, enables reinforcement learning policies to train up to 6 times faster in non-contact scenarios and 40 times faster in contact-rich ones compared to prior simulators, without sacrificing accuracy.

Wenzhe Tong

Jonathan Mi

Xili Yi

Nima Fazeli

Xiaonan Huang

Tensegrity robots, built from rigid struts held together by tension cables, can roll and tumble through rough terrain, but they have had no good way to feel what they touch. This collaboration between the labs of Xiaonan Huang and Nima Fazeli, associate professor of robotics, packs a camera, LED ring, and elastomer shell into each endcap of the robot, then uses a neural network to translate how the shell deforms into full six-axis force and torque measurements. Tested on a 12-kilogram robot, the sensor reliably detects ground contact, and a novel 3D-printed bonding technique keeps the whole design durable, lightweight, modular, and open source.

Chi Zhang

Mingrui Li

Wenzhe Tong

Xiaonan Huang

How should a robot with no legs, spine, or wheels learn to move? By understanding its own body plan. This work, from Xiaonan Huang’s lab and published in IEEE Robotics and Automation Letters, represents a tensegrity robot’s physical topology as a graph and feeds it to a graph neural network policy, which learns coupling among the robot’s struts and cables for faster, more stable training than standard approaches. Policies learned in simulation transferred directly to a physical three-bar tensegrity robot without fine-tuning, achieving stable straight-line tracking and turning in the real world.

Fan Yang

Zixuan Huang

Abhinav Kumar

Sergio Aguilera Marinovic

Soshi Iba

Rana Soltani Zarrin

Dmitry Berenson

When a robot hand fumbles an object, the difference between a minor stumble and a catastrophic failure is whether it can catch the object and get back to work. CADRE, from Dmitry Berenson’s lab and collaborators, is a reinforcement learning framework that catches a falling object while it is still within grasping range, then resets the hand to a configuration ready to resume the original task. By reasoning about finger-object contact through implicit neural descriptors rather than object pose alone, CADRE trains more efficiently, recovers more successfully, and generalizes zero-shot to objects it has never seen.

Challen Enninful Adu

César E. Ramos Chuquiure

Yutong Zhou

Pearl Lin

Ruikai Yang

Bohao Zhang

Shubham Singh

Ram Vasudevan

Generating a smooth, dynamically feasible trajectory for a complex robot in seconds is hard enough when minimizing a single cost function, and real tasks call for all kinds of cost functions and constraints. This work from Ram Vasudevan’s lab extends the Affine Geometric Heat Flow (AGHF) PDE, previously limited to a single cost type and initial guesses that already satisfied constraints, to handle arbitrary objective functions. A new two-phase algorithm lets planning start from constraint-violating initial guesses while still guaranteeing convergence, tested across multiple robotic systems including the Digit humanoid.

Annette M. Masterson

Wonse Jo

Helena C. Sieh

Lionel P. Robert Jr.

Dawn Tilbury

Museums increasingly use robots to entertain and educate visitors, but a single physical robot can only be in one place, playing one role, at a time. This work from the labs of Lionel P. Robert Jr. and Dawn Tilbury pairs a physical tour-guide robot with a projected virtual agent that converses and interacts alongside it, giving visitors the richness of two guides from a single platform. In a study with 30 participants, engagement and overall quality of experience held steady across conditions. Participants said in interviews that they preferred the mixed-agent team.