⚡ electrical engineering

A Study on Dynamic Obstacle Avoidance Using the UR5 Robot Arm

This paper proposes a hierarchical online obstacle avoidance framework for the UR5 robotic arm that integrates offline RRT* planning with APF-based trajectory pre-bending and a two-layer soft-constrained linear MPC, achieving real-time performance, smooth motion, and 100% safety compliance in dynamic environments through independent capsule-based safety gating.

Original authors: 敬业段, 谭飞

Published 2026-08-18

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: 敬业段, 谭飞

Original paper licensed under CC BY 4.0 (https://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

In the bustling factories of tomorrow, robots are no longer confined to caged cages, working in isolation from human hands. They are moving into shared spaces, collaborating alongside people to assemble cars, pack boxes, and handle delicate materials. This shift demands a fundamental change in how these machines think. A robot arm in a static factory can follow a pre-programmed path with perfect precision, but in a dynamic world where humans and objects move unpredictably, that same rigid path becomes a hazard. The core challenge for engineers is to teach a robot to see a moving obstacle, predict where it will be a split second from now, and smoothly steer around it without stopping, jerking, or crashing. This requires a delicate balance: the robot must be fast enough to react in real-time, smart enough to plan a safe route, and gentle enough to keep its movements fluid so it doesn't damage itself or the items it carries.

Researchers at Sichuan University of Science and Engineering have tackled this problem by developing a new, three-layered strategy for a standard six-jointed robot arm known as the UR5. Their approach, detailed in a recent study, moves away from relying on a single, complex algorithm to solve every problem at once. Instead, they broke the task down into three distinct phases: a long-term plan, a mid-term adjustment, and a final safety check. The goal was to create a system that could navigate a workspace filled with moving spheres and boxes, ensuring the robot never collides while keeping its movements smooth and its reaction times fast enough to run on standard industrial controllers.

The first layer of their system happens before the robot even starts moving. The researchers used a method called a rapidly-exploring random tree to map out a safe, collision-free path through the static parts of the room, like fixed pillars or walls. But they didn't stop there. Recognizing that moving obstacles often follow predictable patterns, they added a second step to this offline planning phase. They simulated how a moving object would likely travel and gently bent the robot's planned path in advance to give it extra room. Imagine a driver who, seeing a car merging onto the highway ahead, slightly adjusts their lane position early on to create a comfortable buffer, rather than waiting until the last second to swerve. This "pre-bending" of the path meant the robot didn't have to make sudden, jerky corrections later on, saving its motors from strain and keeping the motion smooth.

Once the robot began its task, the second layer of the system took over. This was the real-time brain of the operation, using a technique called model predictive control. Think of this as a driver constantly looking ahead, checking the road for the next few seconds, and making tiny steering adjustments to stay in the lane. The researchers designed this controller to have two different "zones" of safety. The first zone was a preferred distance, a comfortable buffer where the robot would gently nudge its path to stay away from moving objects. The second zone was a hard safety line, a boundary that the robot could not cross under any circumstances. By treating these two zones differently, the system could be flexible when things were calm but become extremely cautious when an obstacle got too close. This dual-layer approach allowed the robot to maintain a smooth trajectory while still having a robust defense against sudden dangers.

However, the researchers knew that even the best predictions can be wrong. If a moving object suddenly changes direction or speed, the robot's plan might no longer be safe. To address this, they added a third, independent layer that acted as a final gatekeeper. Before the robot executed any command, this layer performed a rapid, geometric check to ensure the arm would not hit anything during the next fraction of a second. If the check failed, the system had a three-step emergency plan: first, it would try to brake; if that wasn't enough, it would push the arm away from the danger; and if all else failed, it would hold its position and sound an alarm. This safety gate ensured that even if the main planning system made a mistake, the robot would not cross the hard safety line.

The team tested this system extensively in computer simulations, pitting it against other common methods used in robotics. In one set of tests, they compared their new method to a standard technique called the artificial potential field, which is known for being simple but often results in jerky, stop-and-go movements. The results were striking. While both methods successfully avoided collisions, the new hierarchical system produced movements that were significantly smoother. The researchers measured the "jerk," or the suddenness of changes in acceleration, and found that their method reduced this value to less than half of what the standard method produced. This means the robot moved with a fluidity that is much gentler on its mechanical parts and better for the delicate objects it might be handling.

Safety was the most critical metric, and the system performed flawlessly in this regard. Across one hundred random scenarios involving moving obstacles, the robot never once breached the hard safety boundary. In contrast, other tested methods that lacked the final safety gate failed completely, crashing into obstacles in every single trial. The researchers also tested how the system handled errors in prediction. They created a scenario where a moving object changed direction sharply, a situation where the robot's prediction of its path would be wrong. Even in these difficult cases, the independent safety gate caught the potential errors and prevented a collision, proving that the system could remain safe even when its predictions were imperfect.

The study also looked at the speed of the system, a crucial factor for real-world use. The entire process of planning, checking, and verifying had to happen within a very tight time window of 24 milliseconds, which is the standard heartbeat of industrial robot controllers. The researchers found that their system met this requirement with ease. The median time to solve the problem was just 7.7 milliseconds, and even in the most complex situations, the system rarely exceeded 16 milliseconds. This left plenty of time for the robot to react to new changes in the environment. Furthermore, the researchers verified that the forces and speeds required to execute these movements were well within the physical limits of the UR5 robot, ensuring that the plan was not just mathematically sound but physically possible.

Through a series of detailed experiments, the researchers demonstrated that separating the tasks of planning, reacting, and verifying creates a more robust and efficient system than trying to do everything at once. By shifting some of the workload to the offline planning phase, they freed up the real-time controller to focus on smooth, precise adjustments. By adding a final safety gate, they ensured that the robot could handle the unexpected without failing. The result is a robot that can work safely alongside humans, moving with a grace and reliability that brings the vision of collaborative robotics closer to reality. The study concludes that while no system can predict the future perfectly, a layered approach that combines foresight with independent verification can create a safety net strong enough to handle the uncertainties of a dynamic world.

Technical Summary: A Study on Dynamic Obstacle Avoidance Using the UR5 Robot Arm

1. Problem Statement

The paper addresses the challenge of generating safe, smooth, and real-time executable motions for a six-degree-of-fore (6-DOF) UR5 robotic arm operating in dynamic, human-robot shared workspaces. Existing approaches face specific limitations:

Sampling-based planning (e.g., RRT):* Typically assumes static environments and struggles to adapt to changing obstacle states, failing to guarantee execution safety in dynamic scenarios.
Reactive methods (e.g., Artificial Potential Fields - APF): While computationally efficient, they lack foresight, rely on current states, and are prone to local optima, trajectory oscillations, and inefficient large evasion maneuvers.
Model Predictive Control (MPC): Offers forward-looking capabilities but suffers from high computational complexity when handling the numerous kinematic and collision constraints of high-DOF arms, making it difficult to solve within strict control cycles (e.g., 24 ms).

The core problem is the trade-off between safety, real-time performance, and dynamic feasibility. Single algorithms struggle to simultaneously satisfy global search, local response, and hard safety constraints, particularly when prediction errors exist.

2. Methodology

The authors propose a hierarchical online obstacle avoidance framework for the UR5 arm, decoupling responsibilities into three distinct layers: Offline Planning, Online Optimization, and Pre-submission Safety Gating.

2.1 System Modeling

Kinematics: The UR5 is modeled using modified Denavit-Hartenberg (DH) parameters.
Geometry: Links are approximated as capsules (uniform radius $L_r = 45$ mm) to ensure continuous, sampling-independent distance calculations for collision avoidance.
Safety Thresholds: Three distance layers are defined:
1. Hard Safety Bottom Line (185 mm): Must not be crossed.
2. Soft Constraint Primary Defense (265 mm): Rises with speed.
3. Preference Spacing (405 mm): Desired operating distance.

2.2 Hierarchical Framework

Offline Layer (Global Planning & Pre-bending):
- RRT:* Generates a collision-free global reference trajectory in a static environment.
- APF Reference Pre-bending: Adjusts the reference trajectory offline based on the nominal (constant-speed) motion trends of dynamic obstacles. This shifts part of the avoidance burden away from the online controller, creating a "pre-bent" path that anticipates obstacles without requiring future knowledge unavailable to the online layer.
Online Layer (Two-Layer Soft-Constrained Linear MPC):
- Prediction: Uses constant-speed extrapolation for dynamic obstacles.
- Optimization: A rolling horizon MPC solves a Quadratic Programming (QP) problem to track the pre-bent reference.
- Two-Layer Soft Constraints: Instead of hard inequalities, the MPC employs two relaxation layers with differentiated weights:
  - Preference Layer (Low weight): Handles routine approaches, gently increasing clearance.
  - Safety Layer (High weight): Activates as the robot approaches the hard safety threshold, prioritizing safety over tracking.
- This structure allows the system to trade tracking performance for safety margin without violating hard limits immediately.
Commit Layer (Pre-submission Safety Gating):
- Independent Verification: Before a command is executed, an independent geometric check verifies the "capsule sweep" clearance between the current configuration and the candidate command over the entire control cycle interval.
- Lipschitz Fast Screening: Uses endpoint bounds to skip full sweeps if the interval is guaranteed safe.
- Three-Level Fallback: If a command fails verification, the system triggers:
  1. Braking path search.
  2. Active evasion via potential field repulsion.
  3. Submission of the configuration with the largest margin (with an alarm) if all else fails.
- This layer acts as a "depth-limiting filter" to prevent intrusion depth and duration, rather than a guaranteed violation eliminator for physically impossible scenarios.

3. Key Contributions

Hierarchical Safety-Aware Framework: A decoupled architecture integrating offline RRT*, APF-assisted pre-bending, and online MPC to separate global path generation, dynamic response, and safety verification.
Two-Layer Soft-Constrained Linear MPC: A novel control structure that differentiates between routine approaches and critical safety threats using weighted relaxation variables, improving smoothness and real-time performance while maintaining safety margins.
Independent Pre-commitment Safety Gate: A geometric verification mechanism combining continuous sweep detection with multi-level fallback, enhancing robustness against prediction errors and model uncertainties.
Systematic Evaluation: Comprehensive validation through baseline comparisons, ablation studies, sensitivity analysis, and Monte Carlo statistical analysis under dynamic conditions.

4. Results

Experiments were conducted in MATLAB using the UR5 kinematic model and OSQP solver.

Safety & Success Rate:
- In the main baseline comparison (Table 3), the method achieved a Task Success Rate of 94.2%.
- In a separate Monte Carlo statistical validation involving 100 random trials with perturbed encounter parameters, the Mission Success Rate (Arrival Without Breach) was 100%, with 0% hard-limit breach rate across all 100 trials.
- No hard-limit breaches occurred across 55 parameter configurations and in high-curvature ambush scenarios.
- The median minimum dynamic clearance was 404.4 mm, with the 5th percentile at 303.1 mm (well above the 185 mm hard threshold).
Performance & Smoothness:
- Joint Jerk: The proposed method reduced joint jerk to 1/2.59 that of the standard APF method (295 rad/s³ vs. 765 rad/s³), significantly improving motion smoothness.
- Real-time Performance: The median online solution time was 7.7 ms, and the 95th percentile was 16.24 ms, meeting the 24 ms control cycle requirement (99.56% compliance).
- Dynamic Feasibility: All joint torques and velocities remained within UR5 rated limits (e.g., peak torque utilization at 78% for the shoulder joint).
Ablation & Sensitivity:
- Two-Layer MPC: Increased minimum clearance by 37–106 mm compared to single-layer structures.
- Pre-bending: Reduced the peak online preference layer relaxation by ~45% in stress scenarios, shifting the avoidance burden to the offline phase.
- Gating: In adversarial trials, the gating mechanism reduced intrusion momentum by 23%, average depth by 25%, and violation duration by 39%.
Prediction vs. Gating:
- Perfect prediction improved the median minimum gap by 11.2 mm compared to constant-speed extrapolation, confirming prediction quality enhances safety margins.
- However, even with severe prediction mismatches (constant-speed extrapolation), the independent gating mechanism maintained zero hard-limit breaches, demonstrating that gating safeguards the bottom line even when prediction is imperfect.

5. Significance and Claims

The paper claims that this hierarchical framework successfully balances safety, smoothness, and real-time performance for dynamic obstacle avoidance on a 6-DOF arm. By decoupling global planning, reference adjustment, and safety verification, the system mitigates the computational burden of online optimization while ensuring robustness against prediction errors.

The authors explicitly state that the pre-submission gate is a depth-limiting filter, not a violation eliminator; it cannot guarantee escape from physically infeasible encounters (e.g., extremely high-speed obstacles sweeping the arm). However, the combination of offline pre-bending, differentiated soft constraints, and independent gating provides a robust engineering solution that maintains hard safety boundaries in 100% of the Monte Carlo statistical trials, outperforming single-layer MPC and reactive baselines in both safety margin and motion smoothness.

A scientific accuracy reviewer checked the draft against the paper and flagged these problems:

Summary claims 100% success rate in 100 random encounters, but the paper reports 94.2% success rate in the main comparison table. (the paper says: "Task Success Rate 94.2%")
Summary claims 100% success rate in 100 random encounters, whereas the paper explicitly states 100/100 for 'Mission Success Rate (Arrival Without Breach)' which is a specific subset, while the general 'Task Success Rate' in Table 3 is 94.2%. (the paper says: "Task Success Rate 94.2%")

Produce a corrected version of the draft. Fix ONLY what the reviewer flagged (verify each point against the paper) and keep everything else — the structure, the level of detail, the terminology — unchanged. Output ONLY the corrected summary.