Taekyung Kim | @tkkim-robot

Taekyung Kim | @tkkim-robot

Taekyung Kim | @tkkim-robot

About

I am a PhD student in Robotics at the University of Michigan, Ann Arbor, working with Prof. Dimitra Panagou. Currently, I am working closely with Toyota Research Institute North America (TRINA), collaborating with Georgios Fainekos and Bardh Hoxha.

My research interests lie in safety-critical planning/control, safe generative robot policy, learning for robot dynamics and control, and model-based reinforcement learning.

Prior to my PhD study, I was a robotics researcher at the AI Autonomy Technology Center of the Agency for Defense Development (ADD) in Korea. I fulfilled my mandatory military service for three years as a Research Officer for National Defense, jointly with the Republic of Korea Air Force.

I received my bachelor's degree (Summa Cum Laude) at the Daegu Gyeongbuk Institute of Science and Technology (DGIST). There, I worked with Prof. Yongseob Lim and Prof. Gyeungho Choi on computer vision and control algorithms for autonomous driving.

Besides, I am a professional SCUBA diver (Open Water SCUBA Instructor @ PADI) and enjoy exploring the underwater world. I love to play tennis and video games. I am also very interested in photography, wine, and traveling.

About Me

Interactive Demo ↓

News

Paper on Safe Model Predictive Diffusion is accepted to ICRA

Paper on Safe Model Predictive Diffusion is accepted to ICRA

Paper on Dynamic Collision Avoidance CBF is accepted to ICRA

Paper on Dynamic Collision Avoidance CBF is accepted to ICRA

Paper on Online CBF for VTOL is accepted to CDC

Paper on Online CBF for VTOL is accepted to CDC

Paper on Safe Crowd Navigation is accepted to IROS

Paper on Safe Crowd Navigation is accepted to IROS

I will be joining TRINA as a Research Resident (Summer 2025)!

I will be joining TRINA as a Research Resident (Summer 2025)!

Paper on Certifiably-Correct Mapping is accepted to RSS

Paper on Certifiably-Correct Mapping is accepted to RSS

Paper on Visibility-Aware RRT is accepted to RA-L

Paper on Visibility-Aware RRT is accepted to RA-L

Paper on adapting CBF parameters is accepted to ICRA

Paper on adapting CBF parameters is accepted to ICRA

I have passed my PhD Qualifying Exam and advanced to candidacy

I have passed my PhD Qualifying Exam and advanced to candidacy

Paper on meta-learning traversability estimation is accepted to IROS

Paper on meta-learning traversability estimation is accepted to IROS

Paper on off-road autonomous driving is accepted to RA-L

Paper on off-road autonomous driving is accepted to RA-L

Certified as Drone Pilot

Certified as Drone Pilot

Paper on active exploration is accepted to RSS

Paper on active exploration is accepted to RSS

Invited Talk: SAEG AI/A Workshop

Invited Talk: SAEG AI/A Workshop

I will be starting my PhD in Michigan Robotics !

I will be starting my PhD in Michigan Robotics !

Paper on traversability estimation is accepted to RA-L

Paper on traversability estimation is accepted to RA-L

Paper on indoor mobile robot navigation is accepted to IEEE Access

Paper on indoor mobile robot navigation is accepted to IEEE Access

Certified as a NVIDIA Jetson AI Ambassador

Certified as a NVIDIA Jetson AI Ambassador

Certified as a Udacity Mentor

Certified as a Udacity Mentor

Paper on sampling-based MPC is accepted to RA-L

Paper on sampling-based MPC is accepted to RA-L

Three research papers are accepted to IROS !

Three research papers are accepted to IROS !

Paper on PINN vehicle dynamics is accepted to IROS

Paper on PINN vehicle dynamics is accepted to IROS

Paper on learning-based optimal controller is accepted to RA-L

Paper on learning-based optimal controller is accepted to RA-L

Paper on curriculum learning is accepted to IEEE Access

Paper on curriculum learning is accepted to IEEE Access

Project Pages

Overview diagram of the Safe Model Predictive Diffusion (Safe MPD) algorithm. The forward process (top) gradually adds noise to an optimal trajectory. The backward (denoising) process (bottom) iteratively refines a random trajectory into an optimal solution. At each step, we generate K candidate trajectories. Our Shielded Rollout mechanism transforms every candidate into a kinodynamically feasible and provably safe trajectory before the Monte Carlo score ascent step.

We present Safe Model Predictive Diffusion (Safe MPD), a training-free diffusion planner for generating provably safe and kinodynamically feasible trajectories. Our algorithm integrates a safety shield directly into the denoising process of a model-based diffusion framework. By enforcing feasibility and safety on every sample throughout the denoising process, our method avoids the common pitfalls of post-processing corrections, such as computational intractability and loss of feasibility. Through a parallelization in GPU, our method achieves sub-second planning times even on challenging, non-convex problems.

Motivation

Diffusion-based planners are powerful but often lack safety guarantees. Common safety strategies like guidance or projection can lead to kinodynamically infeasible paths or be computationally intractable in complex, non-convex environments.

Applying Model-Based Diffusion (MBD) to robotics faces two major hurdles:

Sampling Inefficiency: Trajectories that are both safe and kinodynamically feasible form a thin manifold of Lebesgue measure zero in the sampling distribution, making the sampling extremely inefficient.

Lack of Safety Guarantees: Sampling-based methods and post-processing correction techniques like guidance do not provide formal safety guarantees.

Preview Experiments

[ICRA 2026] Safe Model Predictive Diffusion

Illustration of the DPCBF mechanism in dynamic obstacle avoidance

We propose a Dynamic Parabolic Control Barrier Function (DPCBF) for nonholonomic robots in dynamic obstacle avoidance tasks. By dynamically shaping its safety boundary based on the distance to and relative velocity of an obstacle, our method provides a less conservative safety margin. Specifically, the parabola’s vertex shifts away from the robot’s origin in proportion to the relative distance, creating a more relaxed safety constraint compared to traditional collision-cone or velocity-obstacle approaches while still guaranteeing safety under input constraints. This less restrictive formulation significantly improves the feasibility of the underlying CBF-based quadratic program (QP), particularly in dense environments where other methods often fail.

Motivation

(left) Illustration of the collision cone CBF (C3BF) mechanism. (right) QP easily becomes infeasible with C3BF.

Methods based on collision cones or velocity obstacles (VO) define the unsafe set as a cone in the relative-velocity space. CBFs based on these methods ensure safety by constraining the robot's relative velocity to remain outside of this cone. However, this strategy has fundamental weaknesses:

It's overly conservative: The safety constraint depends only on the angle of the relative velocity, not its magnitude or the actual distance to the obstacle. This means the robot is prohibited from moving toward an obstacle, regardless of its speed or how much collision-free space is available.

It fails in dense environments: When a robot is surrounded by many obstacles, the individual collision cones can overlap and eliminate all feasible control options. This causes the underlying QP to become infeasible, leading to mission failure even when a safe path clearly exists

DPCBF

[ICRA 2026] Dynamic Parabolic CBF

Conceptual illustration of the inner safe set and locally validated CBF parameter.

We present a novel theoretical framework for online adaptation of Control Barrier Function (CBF) parameters, i.e., the class K functions, under input constraints. To this end, we propose the concept of locally validated CBF parameters, where adapting these parameters ensures that the system trajectories remain safe within the finite horizon. (a) A candidate inner safe set defined via an Input Constrained CBF cannot be rendered forward invariant using the given CBF parameter. (b) With locally validated CBF parameters, the trajectory remains within the inner safe set over the finite horizon, ensuring safety for that interval. By adapting the CBF parameters, the corresponding inner safe set is reshaped dynamically, alleviating conservatism by allowing the trajectory to extend beyond a fixed, globally verified inner safe set.

Motivation

Control Barrier Functions (CBFs) are widely used in robotics to ensure system safety. However, finding the CBF parameters (class K functions) is particularly challenging for highly nonlinear and complex systems and becomes even more difficult when input constraints are present. Traditional approaches often rely on manually tuning the parameters of the class K functions for the CBF conditions a priori. The performance of CBF-based controllers is highly sensitive to these fixed parameters, potentially leading to overly conservative behavior (such as deadlock) or safety violations (due to infeasibility).

How to Predict Locally Validated CBF Parameters?

Neural network configurations for CBF parameter adaptation

To adapt CBF parameters online, (a) prior works employ deterministic neural networks that directly output the optimal CBF parameter; however, these approaches do not account for the inherent uncertainty in the predictions, which may lead to catastrophic failures when facing out-of-distribution inputs. (b) In contrast, we propose a probabilistic verification method that, instead of outputting a single set of CBF parameters, takes a candidate set and predicts the distributions of the safety and performance characteristics. This key novelty enables rigorous verification of locally validated CBF parameters despite neural network uncertainty.

Preview Experiments

[CDC 2025] How to Adapt CBFs?

Baseline mapping (b) incorrectly labels obstacle voxels as free space due to odometry drift. Our certified map (c) deflates the reported safe region so that safety is preserved at every frame.

Accurate state estimation and mapping are prerequisites for safe navigation, yet most pipelines assume perfect pose estimates. Incremental drift in vision-based odometry (VIO/SLAM) can cause a map to misclassify obstacles as free space, leading to collisions.

We introduce a certifiably-correct mapping framework that deflates the claimed safe region at each timestep using the covariance of the incremental pose estimate. The deflation guarantees that the map stored in the robot’s body frame remains a subset of the true free space, even while global pose error grows unbounded.

Key points

Provably safe: the shrunken set is always a subset of the true free space.

Two plugins:

How It Works

At every frame we take the odometry pose and its covariance, project that uncertainty onto the current map, and shrink the reported safe region so it remains a subset of the true free space.

Certified-SFC – In SFCs (convex polytopes), each face is moved inward by a covariance-based margin, keeping the corridor entirely safe.

Certified-ESDF – In ESDFs we subtract a voxel-wise safety radius, ensuring the distance map never over-estimates clearance.

[RSS 2025] Certifiably-Correct Mapping

Overview diagram of the Online Adaptive ICCBF algorithm applied to MPC framework.

The Online Adaptive ICCBF algorithm dynamically adapts Input Constrained Control Barrier Function (ICCBF) parameters to optimize performance while ensuring safety for input-constrained nonlinear systems. Our approach leverages a Probabilistic Ensemble Neural Network (PENN) to predict performance and risk metrics, considering both epistemic and aleatoric uncertainties. The algorithm incorporates a two-step verification process using Jensen-Rényi Divergence (JRD) and Distributionally-Robust Conditional Value at Risk (DR-CVaR) to identify valid parameters. By adapting ICCBF parameters online based on the current state and nearby environment, our method optimizes performance while maintaining safety.

Motivation

Control Barrier Functions (CBFs) are widely used in robotics to ensure system safety. However, finding valid CBFs that guarantee persistent safety and feasibility remains an open challenge, especially in systems with input constraints. Traditional approaches often rely on manually tuning the parameters of the class K functions for the CBF conditions a priori. The performance of CBF-based controllers is highly sensitive to these fixed parameters, potentially leading to overly conservative behavior (such as deadlock) or safety violations (due to infeasibility).

Algorithm Breakdown

Data Generation

We initially generate robot trajectories using the CBF-based controller to form the training dataset by varying the robot's initial state, obstacle configurations, and class-K function parameters. The risk level and deadlock time are recorded as the ground truth for prediction. The risk level is computed as the maximum safety loss value during navigation, defined by a safety loss density function that captures the collision risk.

[ICRA 2025] Online Adaptive ICCBF

Main illustration of the Visibility-Aware RRT* algorithm.

The LQR-CBF-Steer function is used as a steering method in RRT* to generate safe trajectories between two nodes in the tree. Our Visibility-Aware RRT* algorithm incorporates two control barrier function (CBF) constraints into the LQR-CBF-Steer function to generate safe and efficient paths for robots with limited sensing capabilities. The collision avoidance CBF ensures the planned path remains collision-free w.r.t. known obstacles, while the novel visibility CBF guarantees the robot stays within locally collision-free regions, enabling timely detection and avoidance of unknown obstacles. These CBF constraints serve as termination criteria during the steering process, ensuring that the generated paths are both collision-free and visibility-aware.

Motivation

Collision with a hidden obstacle due to limited sensing capabilities and a perception-agnostic global planner.

Safe autonomous navigation in unknown environments is challenging for robots with limited sensing capabilities. The effectiveness of safety-critical control techniques, such as Control Barrier Functions (CBFs), relies on the assumption of complete environmental knowledge. The video demonstrates that a CBF-based safety-critical controller (CBF-QP) becomes infeasible and collides with a hidden obstacle while tracking a reference path generated by the standard LQR-RRT*, which is agnostic to the robot's sensing limitations. This illustrates that if the global planner does not account for the robot's limited perception, the theoretical safety guarantees may not hold for the entire navigation stack, leading to potentially hazardous situations. This highlights the need for a planning algorithm that explicitly considers the robot's limited sensing capabilities to ensure safe navigation in partially unknown environments.

Preview Experiments

Visualization of the global planning results generated by the Visibility-Aware RRT*, without and with incorporating visibility CBF. The blue and yellow squares represent the start and goal positions. The black circles represent the known obstacles. The red line depict the final reference path. The shaded areas in gray represent the local collision-free set that the robot will sense while following the reference path.

[RA-L 2025] Visibility-Aware RRT*

Overview diagram of our unified model-based reinforcement learning framework with dynamics learning. In exploration phase, a parallelized ensemble neural network serves as the robot dynamics and outputs the estimated posterior distribution of the next state. To enable active exploration, we quantify epistemic uncertainty by measuring the ensemble disagreement via Jensen-Rényi Divergence. In deployment phase, the neural network dynamics trained during the active exploration phase is applied directly to perform uncertainty-aware control. We transfer the neural network dynamics for uncertainty-aware deployment with minimal modification.

Video

Main Supplmentary Video

Real-World Demo

We successfully transferred our algorithm to the real-world settings for uncertainty-aware deployment tasks. These are our initial real-world experiments using zero-shot transfer, without considering sim-to-real techniques such as domain adaptation or domain randomization. We integrated our algorithm with global path planning and online traversability map generation using a LiDAR sensor. These experiments were conducted on our off-road testbeds. See ‘Future Work’ section below for more details.

[RSS 2023] Bridging Active Exploration and Uncertainty-Aware Deployment

Direct to the project page

Hojin Lee*, Taekyung Kim*, Jungwi Mun, Wonsuk Lee (* equal contribution) AI and Autonomy Technology Center, Agency for Defense Development arXiv | video

https://sites.google.com/view/terrainawarekinodyn

[RA-L 2023] Learning Terrain-Aware Kinodynamic Model for Off-Road Driving

A simplified representation of the MPPI algorithm during each optimization iteration. For clarity, we only visualize one sampled trajectory (in green). (a) Amount of changes between previously computed control sequence and the next control sequence (along the “i-axis”). (b) Amount of changes in control values during MPPI rollouts (along the “t-axis”), which are hard to be minimized by the MPPI baseline. Such chattering in control input becomes more prominent in cases where the environment changes rapidly, possibly even causing the MPPI to diverge.

To address this issue, we propose the Smooth MPPI algorithm that seamlessly combines MPPI with an input-lifting strategy.

Frequently Asked Questions

Q: How to tune Δt in Smooth MPPI?

A: The sampling time Δt affects the covariance of the control input derivative significantly. Here is the suggested way to tune the parameters:

Tune the covariance of the control input with normal MPPI

Implement the Smooth MPPI by setting Δt = 1 (even if the sampling time is not 1 sec)

Fine tune the covariance of the control input derivative slightly based on the tuned covariance of the normal control input from step 1. 

ex) If the covariance in MPPI was Diag(1,1), the covariance of the SMPPI can be tuned with the range of Diag(0.1, 0.1) to Diag(1,1). 

It is fine to set ∆t to be the actual sampling time, but in this case, it’s hard to tune the covariance due to numerical issue. You have to increase it by a large magnitude.

[RA-L 2022] Smooth MPPI

Projects

Off-Road Autonomous Driving

Uncertainty-Aware Active Exploration

Traversability Estimation

Learning-Based Vehicle Model and Control

Contact

My laptop background image