Institution profile

Robotics and AI Institute

Academic institution
Research library7linked papers
Opportunities0open roles
Selected work

Representative Papers

Bicycle Acrobatics with Reinforcement Learning

Aug 01, 2026

This work addresses the challenge of achieving high-agility stunts on bicycle robots, which are inherently difficult due to their simple structure, underactuation, and nonholonomic constraints. The study presents the first application of reinforcement learning to this platform, introducing a multi-task RL framework that integrates waypoint tracking, posture control, payload following, and motion imitation. A state-triggered policy coordinator is designed to seamlessly compose diverse dynamic maneuvers. On the authors’ custom-built UMV bicycle robot, the approach successfully executes complex stunts—including continuous jumps, backflips, wheelies, and three-point turns—enabling it to clear 1-meter-high obstacles, perform over 15 consecutive jumps, and execute action sequences exceeding 20 steps. Both simulation and real-world experiments demonstrate the method’s robustness, generalization capability, and its ability to surpass the agility limitations typical of conventional wheeled platforms.

0 citationsRead paper

Show, Don't Tell: Detecting Novel Objects by Watching Human Videos

Mar 13, 2026

Existing robotic systems struggle to rapidly recognize out-of-distribution novel objects from human demonstrations: closed-set detectors fail, while open-set approaches rely on cumbersome language prompts. This work proposes a “show-don’t-tell” paradigm that leverages self-supervised learning to automatically construct training data and generate supervision signals directly from human manipulation videos, enabling the customization of novel object detectors without any textual descriptions. The method achieves end-to-end rapid adaptation of object detectors and significantly outperforms current state-of-the-art techniques on real robotic platforms, substantially improving task completion rates.

0 citationsRead paper

Learning Smooth Time-Varying Linear Policies with an Action Jacobian Penalty

Feb 20, 2026

This work addresses the challenge that real-world robots and humans struggle to reproduce the high-frequency, unnatural control signals commonly generated by reinforcement learning policies. To this end, the authors propose a task-agnostic action Jacobian penalty that explicitly constrains the policy’s sensitivity to state variations, coupled with a computationally efficient Linear Policy Network (LPN) to enable rapid training and inference. The resulting approach significantly enhances policy smoothness and convergence speed, successfully generating naturalistic motions—including backflips and complex parkour maneuvers—while demonstrating physical feasibility on a quadrupedal robot equipped with a manipulator arm for dynamic tasks.

0 citationsRead paper
Recent publications

Latest Papers

Bicycle Acrobatics with Reinforcement Learning

Aug 01, 2026

This work addresses the challenge of achieving high-agility stunts on bicycle robots, which are inherently difficult due to their simple structure, underactuation, and nonholonomic constraints. The study presents the first application of reinforcement learning to this platform, introducing a multi-task RL framework that integrates waypoint tracking, posture control, payload following, and motion imitation. A state-triggered policy coordinator is designed to seamlessly compose diverse dynamic maneuvers. On the authors’ custom-built UMV bicycle robot, the approach successfully executes complex stunts—including continuous jumps, backflips, wheelies, and three-point turns—enabling it to clear 1-meter-high obstacles, perform over 15 consecutive jumps, and execute action sequences exceeding 20 steps. Both simulation and real-world experiments demonstrate the method’s robustness, generalization capability, and its ability to surpass the agility limitations typical of conventional wheeled platforms.

0 citationsRead paper

Show, Don't Tell: Detecting Novel Objects by Watching Human Videos

Mar 13, 2026

Existing robotic systems struggle to rapidly recognize out-of-distribution novel objects from human demonstrations: closed-set detectors fail, while open-set approaches rely on cumbersome language prompts. This work proposes a “show-don’t-tell” paradigm that leverages self-supervised learning to automatically construct training data and generate supervision signals directly from human manipulation videos, enabling the customization of novel object detectors without any textual descriptions. The method achieves end-to-end rapid adaptation of object detectors and significantly outperforms current state-of-the-art techniques on real robotic platforms, substantially improving task completion rates.

0 citationsRead paper

Learning Smooth Time-Varying Linear Policies with an Action Jacobian Penalty

Feb 20, 2026

This work addresses the challenge that real-world robots and humans struggle to reproduce the high-frequency, unnatural control signals commonly generated by reinforcement learning policies. To this end, the authors propose a task-agnostic action Jacobian penalty that explicitly constrains the policy’s sensitivity to state variations, coupled with a computationally efficient Linear Policy Network (LPN) to enable rapid training and inference. The resulting approach significantly enhances policy smoothness and convergence speed, successfully generating naturalistic motions—including backflips and complex parkour maneuvers—while demonstrating physical feasibility on a quadrupedal robot equipped with a manipulator arm for dynamic tasks.

0 citationsRead paper