💻 computer science

BIG-CBF: Behavior-Imagination-Guided Control Barrier Function with Shared Uncertainty for Mobile Robot Navigation

This paper introduces BIG-CBF, a two-rate navigation architecture that integrates behavior imagination with shared uncertainty modeling to guide Control Barrier Functions, thereby achieving near-perfect task success and minimizing safety filter interventions in mobile robot navigation.

Original authors: Shibo Li, Zhongcheng Wang, Jiahe Cao, Jianhua Yang, Ke Wu

Published 2026-09-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shibo Li, Zhongcheng Wang, Jiahe Cao, Jianhua Yang, Ke Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot trying to walk through a crowded room. Its primary job is to avoid bumping into people or furniture, a task usually handled by a safety system that acts like an invisible shield. This shield is mathematically designed to stop the robot the instant it gets too close to an obstacle, ensuring it never crashes. However, this safety shield has a blind spot. It is excellent at saying "stop" or "slow down," but it is often terrible at deciding which way to go next. When a robot faces a complex situation, like a narrow hallway with people moving in both directions, the safety shield might keep the robot safe by freezing it in place, preventing a collision but also preventing progress. The robot becomes stuck, safe but useless, unable to figure out whether to step left, step right, or wait. This is the core problem researchers in robotics face: how to keep a machine safe without making it so cautious that it never finishes its job.

To solve this, a team of researchers has developed a new navigation system called BIG-CBF. The name stands for Behavior-Imagination-Guided Control Barrier Function, but the idea behind it is surprisingly straightforward. Instead of just reacting to danger the moment it appears, the robot now pauses for a split second to "imagine" a few different ways it could move. It tests six specific strategies in its mind, such as "walk straight," "step left to pass," "step right to pass," or "wait for the crowd to clear." The system then picks the single best strategy that keeps it moving toward its goal while staying safe. Once a strategy is chosen, the robot executes it, but it still keeps the original safety shield active as a final backup. The key innovation is that the robot uses the same understanding of uncertainty for both the mental test and the physical movement. In the real world, sensors are not perfect, and there is always a tiny delay between seeing something and moving. By acknowledging these delays and errors in both the planning stage and the execution stage, the robot avoids a common trap where a move looks good on paper but fails in reality because the safety system suddenly panics.

The researchers tested this new approach in a rigorous series of simulations involving 3,600 different episodes across nine distinct scenarios. These scenarios ranged from navigating through a dense crowd of moving people to squeezing through narrow gaps between static obstacles. In every single test, the robot was equipped with the same underlying safety shield, but the way it chose its path varied. The results were striking. The new BIG-CBF system successfully completed 99.78% of the tasks, a rate significantly higher than other advanced methods. More importantly, it required far less intervention from the safety shield. In previous systems, the safety shield had to constantly correct the robot's commands, often forcing it to stop or swerve unexpectedly. With BIG-CBF, the robot's chosen path was so well-aligned with safety requirements that the shield rarely had to step in. This means the robot moved more smoothly and efficiently, spending less time frozen in indecision.

To ensure these results were not just a result of the computer, the team took the system to the real world. They installed the software on a physical, omnidirectional robot equipped with a powerful onboard computer. The robot was put through fifteen real-world trials in three challenging environments: a scenario requiring a deliberate sidestep, a room filled with static clutter, and a path crossing with pedestrians. In every single one of these fifteen runs, the robot reached its destination without making contact with any object or person. While other methods struggled, with some failing to complete even a fraction of their trials, BIG-CBF succeeded every time. The data showed that by sharing a consistent view of uncertainty between the planning and execution layers, the robot reduced the energy spent on safety corrections by a significant margin. It did not just avoid crashes; it maintained a steady, purposeful flow of movement.

The study also revealed a subtle trade-off. Because the system is designed to be robust against uncertainty, it sometimes chooses to be slightly more conservative than absolutely necessary, such as waiting a moment longer before moving. This added a small amount of time to the total journey, but it was a deliberate choice to ensure safety rather than a failure to find a path. The researchers found that this small delay was a fair price to pay for the dramatic reduction in safety system interventions and the elimination of the "stuck" behavior that plagues other robots. The system proved that a robot does not need to be a supercomputer to think ahead; it simply needs a structured way to imagine a few simple options and commit to the best one.

This work demonstrates that the future of safe robot navigation lies not just in better sensors or faster computers, but in better decision-making logic. By separating the choice of a general strategy from the split-second execution of that strategy, and by ensuring both layers understand the same limitations of the real world, robots can navigate complex, dynamic spaces with a level of confidence that was previously out of reach. The results suggest that with the right framework, autonomous machines can be both rigorously safe and reliably productive, moving through our world without hesitation or unnecessary stops.

Technical Summary: BIG-CBF

Problem Statement
Control Barrier Functions (CBFs) provide a mathematically rigorous framework for enforcing local collision avoidance in mobile robots, typically acting as safety filters that modify commands to satisfy forward-invariance conditions. However, standard minimum-intervention CBF filters suffer from a "liveness gap." While they ensure safety, they lack task-level maneuver awareness. In geometrically ambiguous environments (e.g., symmetric obstacles or narrow passages), a CBF may suppress unsafe approach components without establishing a persistent lateral preference, causing the robot to remain safe but become trapped in a stalled state.

Furthermore, practical deployment faces a "cross-layer mismatch" under uncertainty. Upstream maneuver selection layers often evaluate candidates using nominal geometry, while downstream safety filters account for uncertainty (e.g., perception noise, execution delays, zero-order-hold effects). This inconsistency can lead to a situation where a maneuver deemed feasible during planning repeatedly triggers corrective interventions during execution, reducing efficiency and increasing the burden on the safety filter.

Methodology: BIG-CBF Architecture
The paper proposes BIG-CBF (Behavior-Imagination-Guided Control Barrier Function with Shared Uncertainty), a two-rate navigation architecture designed to resolve maneuver ambiguity and align uncertainty semantics across control layers.

Two-Rate Architecture:
- Low-Rate Behavior Imagination ( $T_B = 0.5s$ ): A lightweight layer evaluates a compact library of six closed-loop feedback behaviors: GOAL, PASS-L, PASS-R, SIDE-L, SIDE-R, and YIELD. Instead of replanning open-loop trajectories, this layer selects a single behavior mode ( $b^*$ ). The selected policy continuously regenerates state-dependent local subgoals and input commands based on the latest measurements.
- High-Rate Execution ( $T_C = 0.05s$ ): The selected behavior generates a nominal command ( $u_{in}$ ), which is then filtered by a hard CBF to produce the final safe command ( $u_{safe}$ ).
Shared Uncertainty Representation:
To eliminate the planning-execution mismatch, BIG-CBF constructs a common relative-position uncertainty envelope used by both the imagination and execution layers. This envelope aggregates four specific uncertainty sources:
- Relative-motion delay (perception-to-actuation).
- Obstacle prediction uncertainty (covariance growth).
- Robot execution residuals (bounded errors).
- Zero-order-hold (ZOH) motion between samples.
The imagination layer uses this uncertainty to analytically evaluate CBF compatibility and score behaviors without solving a quadratic program (QP). The execution layer uses the same uncertainty semantics in the hard CBF constraint. This ensures that maneuvers selected upstream are consistent with the safety constraints encountered downstream.
Scoring and Selection:
Behaviors are ranked based on a cost function ( $J_b$ ) that balances:
- CBF Compatibility: An analytic penalty for potential barrier violations over the horizon.
- Task Progress: Distance to the goal.
- Freezing: Penalizing commands that result in near-zero velocity.
- Smoothness: Penalizing high control effort or switching.
- Switching: Penalizing unnecessary changes in behavior mode.
A candidate is rejected if the robust barrier test fails; otherwise, the lowest-cost accepted candidate is selected, subject to dwell-time and switching-improvement thresholds.

Key Contributions

Lightweight Behavior Imagination: Introduces a maneuver-selection architecture that evaluates a compact library of closed-loop behaviors over a finite horizon. It resolves discrete passing and yielding choices without dense trajectory optimization or replaying open-loop trajectories.
Cross-Layer Shared-Uncertainty Evaluation: Formulates a unified relative-position uncertainty representation that aligns behavior-level evaluation with execution-level safety constraints. This reduces the need for downstream correction by ensuring upstream candidates are evaluated under the same uncertainty semantics as the execution filter.
Extensive Empirical Validation: Validates the approach through 3,600 simulation episodes across nine diverse scenarios and 120 supervised trials (8 methods × 3 scenarios × 5 repeats) on an omnidirectional robot equipped with a Jetson Orin Nano.

Results

Simulation Performance: Across 3,600 episodes, BIG-CBF achieved a 99.78% task success rate, outperforming baselines including CBF-only, Backup-CBF, PL-CBF, APF, DWA, and MPPI. It substantially reduced the burden on the downstream safety filter, with a mean CBF intervention energy (ECBF) of 0.02 compared to 0.83 for DWA+CBF and 0.45 for MPPI+CBF. The CBF activation ratio dropped to 9.10% (vs. 30.24% for DWA).
Ablation Study (Shared Uncertainty): Comparing BIG-CBF with a variant lacking shared uncertainty (BIG-CBF-NoShared) revealed that shared uncertainty reduced the CBF activation ratio from 15.64% to 9.10% and intervention energy from 0.10 to 0.02. This came with a modest trade-off in completion time (11.72s vs. 11.21s), attributed to more conservative but consistent maneuver selection.
Hardware Experiments: On a physical Rosmaster X3 robot, BIG-CBF completed 15/15 supervised trials (3 scenarios × 5 repeats) without contact events. It demonstrated onboard computational feasibility, with a planner latency of 141.0 ms (within the 0.5s budget) and a CBF solve time of 3.86 ms. The hardware results confirmed that shared uncertainty reduced CBF intervention energy by 0.34 and activation rate by 12.4 percentage points compared to the non-shared variant.

Significance and Claims
The paper claims that BIG-CBF successfully bridges the gap between local safety filtering and task-level liveness. By decoupling maneuver selection from high-rate safety filtering and enforcing consistency in uncertainty modeling across layers, the framework prevents robots from stalling in ambiguous geometries while minimizing the frequency and magnitude of safety filter interventions. The authors assert that this approach offers a computationally efficient alternative to heavy trajectory sampling methods (like MPPI) while providing superior robustness and task completion rates compared to standard reactive or library-based CBF approaches. The work demonstrates that lightweight behavior imagination, when coupled with shared uncertainty, can improve task liveness without compromising the safety guarantees of hard CBF execution.

Technical Summary: BIG-CBF

More like this