Post

ICRA 2026 BARN Challenge: NMPC Navigation on the Jackal

Team Robo's NMPC navigation stack (acados, Clearpath Jackal) for the ICRA 2026 BARN Challenge: 3rd place in the physical finals in Vienna.

ICRA 2026 BARN Challenge: NMPC Navigation on the Jackal

The BARN (Benchmark Autonomous Robot Navigation) Challenge asks teams to drive a Clearpath Jackal through dense, narrow obstacle courses as fast as possible without touching anything. The robot has a single 270° 2D LiDAR and a 2 m/s top speed. I led Team Robo from NTU (8 members) in the 2026 edition, held at ICRA 2026 in Vienna.

StageResult
Simulation (17 teams)6th, score 0.4515, invited to the physical finals
Physical finals, Vienna, 2–4 June 20263rd place, 3 of 9 trials successful
Dynamic-environment benchmark1st all-time

Running the Jackal in the physical finals Physical finals at the Vienna Congress & Convention Center. The course is built from cardboard boxes and changes between rounds.

Two runs from the physical final

Left: trial 1. Right: trial 2. The Jackal runs fully autonomously on its 2D LiDAR through the cardboard course.

Award ceremony for the Fifth BARN Challenge at ICRA 2026 Award ceremony at ICRA 2026, Vienna: Team Robo from NTU in third place.

Certificate for 3rd place in the Fifth BARN Challenge The IEEE Robotics and Automation Society certificate for 3rd place.

Approach: one NMPC, several modes

Most BARN stacks are tuned for static clutter. We built a nonlinear model predictive controller (NMPC) that can handle static walls and moving obstacles in the same optimisation, solved in real time with acados. Many times a second it plans a short stretch of motion ahead, applies the first step, and plans again.

A Jackal dodging moving obstacles in a DynaBARN-style course The idea in one picture: moving obstacles are tracked, their future positions predicted, and the planned path bends around where they will be, not where they are.

The robot model

The Jackal is a differential-drive robot. We model its two wheel speeds directly, so the controller respects how quickly the real wheels can speed up and slow down:

\[\begin{gathered} v = \frac{v_r + v_l}{2}, \qquad \omega = \frac{v_r - v_l}{L} \\[6pt] \dot x = v\cos\theta,\quad \dot y = v\sin\theta,\quad \dot\theta = \omega \end{gathered}\]

The optimisation, in words

\[\begin{gathered} \min_{\text{controls}}\; \underbrace{\text{tracking error}}_{\text{stay on the path}} + \underbrace{\text{control effort}}_{\text{move smoothly}} \\[8pt] \text{subject to}\quad \underbrace{\text{robot limits}}_{\text{speed, turning}},\;\; \underbrace{\text{clearance}}_{\text{now and predicted}} \end{gathered}\]

The clearance constraints are soft: in a gap tighter than the safety margin, the solver may give up a little clearance at a high cost instead of failing outright. In a course as tight as BARN, that difference matters.

Modes: same solver, different priorities

The repository is called mode-switching MPC because the controller changes what it prioritises depending on the situation. The optimisation stays the same; each mode adjusts what it is asked to do.

flowchart LR
  S{"Moving obstacle<br/>nearby?"} -- yes --> D[Dynamic]
  S -- no --> G{"Goal close<br/>and clear?"}
  G -- yes --> R[Rush to goal]
  G -- no --> W{"Walls close?"}
  W -- yes --> T[Static]
  W -- no --> N[Normal]
  • Normal: open space, go fast.
  • Static: tight walls, move carefully.
  • Dynamic: something is moving nearby, plan around where it is heading.
  • Rush to goal: the finish is close and clear.
  • When the path doubles back, the robot can also turn on the spot or reverse out instead of attempting a wide turn.

Tracking moving obstacles

The moving obstacles come from our obstacle tracker, built on Mateusz Przybyła’s obstacle_detector. It turns the laser scan into shapes and follows them over time to estimate how fast each one is moving.

A raw tracker reports everything, including walls that only seem to move because of sensor noise. We added a filter so that only genuinely moving obstacles reach the controller, which stops the robot from swerving around phantoms.

Raw tracker output versus what reaches the controller Left: the raw tracker sees “motion” in wall noise. Right: after our filter, only the real moving obstacle is passed on.

Simulation vs. the physical finals

The work lives in two branches of the repository:

  • v3.0, simulation: the full stack including moving-obstacle tracking, used for the dynamic benchmark.
  • main, physical finals: the version that ran in Vienna. The 2026 courses were static, so tracking was switched off. It is packaged as a standard move_base plugin and hardened for real sensor noise.

Team at the physical finals Setting up between trials. Each team gets a fixed time slot per course.

Context from the finals

  • Simulation and physical rankings differed. Only four of the seven invited teams competed in Vienna, and the order changed: we placed 6th in simulation and 3rd on the real robot.
  • The 2026 courses were static only. Unlike earlier editions, there was no dynamic-obstacle component. Our dynamic handling went unused in the finals, but it is what ranks us first all-time on the dynamic-environment benchmark.
  • Every team in the finals used classical methods. For the second year in a row, no finalist used machine learning in their navigation stack.

Our approach is described in the organisers’ article on the competition: Autonomous Ground Navigation in Highly Constrained Spaces: Lessons Learned from the Fifth BARN Challenge at ICRA 2026, submitted to IEEE Robotics and Automation Magazine.

My role

  • Led the 8-member team through the simulation qualifier and the physical finals.
  • Wrote most of the mode-switching MPC planner (C++, acados, ROS move_base plugin).
  • Co-developed the laser-based dynamic obstacle tracker, including the filter that keeps only genuinely moving obstacles.
  • Also built and tested alternative planners we compared against: a D* Lite global planner, a visibility-graph global planner, and DDP-based local planning.
This post is licensed under CC BY 4.0 by the author.