⚡ electrical engineering

Reactive Online Motion Generation for Redundant Manipulators with Adaptive Manipulability Weighting

This paper proposes an online motion generation method for redundant manipulators that dynamically adjusts the priority between goal attainment and manipulability improvement via an adaptive weighting scheme and a reference ellipsoid, enabling the robot to avoid singularities and reach targets efficiently without a fixed reference trajectory.

Original authors: Ryuji Nakagawa, Ryosuke Tasaki

Published 2026-10-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryuji Nakagawa, Ryosuke Tasaki

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

Robots with many joints, like the human arm, possess a remarkable flexibility that allows them to reach the same point in space with their hand in countless different ways. This flexibility is a double-edged sword. While it offers freedom, it also creates a hidden danger: certain positions can leave the robot stiff, unable to move in specific directions without straining its motors or losing control entirely. Engineers call these precarious positions singularities. To keep a robot working smoothly, it must constantly avoid these dead ends while still getting its job done. The challenge lies in balancing two competing needs: reaching a destination quickly and maintaining a posture that keeps the robot agile and ready for whatever comes next.

For years, researchers have tried to solve this by pre-calculating safe paths or by using fixed rules that treat every movement the same way. However, these methods often struggle when the starting position of the robot changes or when the environment is unpredictable. A new study by Ryuji Nakagawa and Ryosuke Tasaki at Aoyama Gakuin University proposes a different approach. Instead of following a rigid, pre-drawn map, their method allows the robot to decide its own path in real time, constantly adjusting its priorities based on how "healthy" its current posture feels. The researchers developed a system that lets a robot move toward a goal while simultaneously checking if it is drifting toward a stiff, awkward position, and if so, gently steering itself away without losing sight of the target.

The core of this new system is a dynamic way of thinking about space and distance. Imagine the robot is given a budget for how far it can travel from a straight line connecting its start point to its goal. If the robot is in a comfortable, flexible position, it spends very little of this budget, staying close to the direct path. But if the robot senses it is entering a stiff configuration, the system uses this budget to allow the robot to take a wider, curved detour to regain its flexibility. This budget is not a fixed number; it shrinks as the robot moves, ensuring that the robot eventually runs out of room for detours and is forced to focus entirely on reaching the goal.

To make these decisions, the researchers created a mathematical tool that acts like a compass for flexibility. At every fraction of a second, the robot calculates which direction would make it most agile. It then looks at the remaining "budget" of distance it can afford to travel and finds the best spot within that limit to move toward. This spot is not chosen by guessing or searching through thousands of options, which would be too slow for a real-time system. Instead, the robot calculates the perfect point instantly using a geometric shape that shrinks as the robot gets closer to its goal. This allows the robot to react instantly to its own changing state, prioritizing flexibility when it is low and speed when it is high.

The researchers tested this idea extensively using a seven-jointed robot arm in computer simulations. They set up three different starting situations: one where the robot began in a comfortable position, one where it started in a very awkward, stiff pose, and one where it was somewhere in between. In every case, the robot successfully reached its destination. When the robot started in a stiff position, the system immediately allowed it to take a longer, curved path to fix its posture before heading to the goal. When the robot started in a good position, it stayed close to the straight line, avoiding unnecessary detours. The system proved that by adjusting its priorities on the fly, the robot could avoid getting stuck in bad positions without needing a different set of rules for every new starting point.

To prove this worked in the real world, the team attached a glass of liquid to the robot's hand and asked it to move the glass from one spot to another without spilling. This required the robot to keep the glass perfectly upright while moving. In one test, the robot started in a position where moving in a straight line would have caused it to become stiff and potentially drop the glass. The new system successfully guided the robot on a curved path that kept the glass upright and the robot flexible, all while moving at a speed of 100 times per second. The computer calculations took less than one millisecond each time, proving the method is fast enough for real-time control.

The study confirms that a robot does not need a pre-planned map to move safely and efficiently. By constantly evaluating its own flexibility and using a shrinking budget of allowable detours, the robot can navigate complex situations on its own. The researchers found that this adaptive approach outperformed older methods that used fixed rules, particularly when the robot started in difficult positions. While the system does not guarantee a perfect path in every possible scenario, it demonstrated a robust ability to balance the need to reach a goal with the need to stay agile. This work suggests a future where robots can handle unpredictable tasks with a level of adaptability that mimics the natural fluidity of human movement, moving not just to a destination, but through the world with awareness of their own physical limits.

Technical Summary: Reactive Online Motion Generation for Redundant Manipulators with Adaptive Manipulability Weighting

Problem Statement
For redundant manipulators, maintaining kinematic capability (manipulability) is critical to avoid singularities where motion directions become restricted and joint velocities may spike. While existing methods address manipulability, they face limitations: offline capability maps are costly to generate and may not reflect the current inverse-kinematics branch; control-level null-space projection improves manipulability only within the redundancy remaining after the primary task, without altering the end-effector path; and planning-level methods often require pre-defined trajectories or complex search trees. Furthermore, balancing the priority between goal attainment and manipulability improvement is difficult when fixed weights are used, as a constant weight may cause unnecessary detours when manipulability is high or fail to provide sufficient guidance when it is low. This study addresses the need for a reactive, online motion generation method that dynamically adjusts the priority between improving manipulability and reaching a goal without a pre-defined intermediate trajectory.

Methodology
The proposed method generates reference velocities in real-time based on the current state, integrating them into a velocity-level Quadratic Program (QP) subject to joint constraints. The core components are:

Path-Length Budget and Reference Ellipsoid:
A path-length budget ( $L_{max}$ ) is defined based on the straight-line distance between the start and goal ( $L_{min}$ ) and an allowable detour ratio ( $r_{detour}$ ). As the end-effector moves, the cumulative path length ( $L(t)$ ) is subtracted from $L_{max}$ to determine the remaining budget ( $B(t)$ ). This budget defines a 3D reference ellipsoid centered between the current position and the goal. The ellipsoid represents the spatial region from which a manipulability reference position can be selected without exceeding the total allowable path length.
Analytical Reference Position Generation:
Instead of searching candidate points, the method analytically computes a manipulability reference position ( $x_m$ ) that maximizes the predicted manipulability within the reference ellipsoid. Using the local translational manipulability gradient ( $g(t)$ ) derived from the current configuration, the reference position is identified as the support point of the ellipsoid in the direction of the gradient. This is achieved via a closed-form solution involving the shape matrix of the ellipsoid, eliminating the need for discrete searches or iterative optimization.
Adaptive Manipulability Weighting:
An adaptive weight ( $w_m$ ) is calculated to determine the priority of manipulability improvement relative to goal attainment. This weight is derived from the relative predicted improvement at the reference position, normalized by the current manipulability and the reference distance.
- When the current manipulability is low and significant improvement is predicted, the weight increases, prioritizing the manipulability direction.
- When manipulability is already high or the gradient is small (near a local maximum), the weight decreases, prioritizing goal attainment.
- This mechanism avoids the need for global maximum manipulability values and prevents the weight from being overly dependent on the arbitrary size of the reference ellipsoid.
Potential-Based Reference Velocities:
Two task-space reference velocities are generated: one for the goal ( $v_g$ ) and one for the manipulability reference ( $v_m$ ). Both follow a distance-dependent law similar to an attractive potential, decaying as the end-effector approaches their respective targets. The manipulability velocity is further attenuated near the goal to prevent excessive deviation as the target is reached.
Velocity-Level Quadratic Program:
The final joint velocity is computed by solving a convex QP that minimizes the error between the actual joint velocity and the weighted sum of the goal velocity, the manipulability velocity, and a null-space posture improvement term. The adaptive weight $w_m$ dynamically balances the first two terms. Joint position and velocity limits are enforced as constraints.

Key Contributions
The paper identifies two primary methodological contributions:

Analytical Reference Position Generation: A method to generate a manipulability reference position that accounts for path-length growth without a candidate-point search. This is based on a reference ellipsoid that contracts as the cumulative path length increases and the local manipulability gradient.
Adaptive Weight Determination: A method to adaptively determine the priority of manipulability improvement based on a relative assessment of the predicted improvement over the current value. This avoids using global maximum manipulability in the weight calculation, allowing the system to adjust priorities based on the immediate kinematic state.

Results
The method was evaluated through simulations with a 7-DOF KUKA LBR iiwa 7 R800 manipulator and real-robot transport experiments.

Simulation (Scenario Analysis):
- Detour Ratio: Increasing the allowable detour ratio ( $r_{detour}$ ) allowed for greater manipulability improvement but yielded diminishing returns beyond a ratio of 2.5.
- Adaptive vs. Fixed Weights: In scenarios with low initial manipulability, the adaptive weight successfully prioritized improvement, preventing the manipulability from dropping below initial levels, whereas fixed weights often failed to reach the goal or required excessive detours. In high-manipulability scenarios, the adaptive weight limited unnecessary path modifications, achieving goal attainment faster than high fixed weights while maintaining better manipulability than low fixed weights. The adaptive method used a single set of design parameters across all scenarios, reducing reliance on scenario-specific tuning.
Real-Robot Experiments (Transport with Fixed Orientation):
- The method successfully generated paths at 100 Hz while maintaining a fixed end-effector orientation (transporting a liquid-filled glass).
- In a standard scenario, the method achieved higher mean and minimum manipulability compared to a trajectory-tracking method (Jin et al.) and a deviation-allowed method (Dufour and Suleiman), even when the latter was tuned to match the path length.
- In a near-singularity scenario, the method improved manipulability without falling below the initial value, successfully avoiding the singular configuration that a straight-line trajectory would have traversed.
- Computation times averaged under 0.5 ms per cycle, confirming real-time feasibility.

Significance and Claims
The authors claim that this reactive online motion generation method effectively balances manipulability improvement and goal attainment without requiring a pre-defined reference trajectory. By analytically generating a reference position within a dynamically updated ellipsoid and using an adaptive weight based on local state predictions, the method reduces the need for manual tuning of manipulability weights. The study demonstrates that the approach can suppress decreases in manipulability during motion and reach goals from various initial states using a common parameter set.

The paper explicitly notes its limitations: the method does not guarantee global optimality, goal attainment from arbitrary initial states, or avoidance of local stagnation. The path-length budget defines an admissible region for selection but does not strictly guarantee the actual path length will not exceed the budget due to discrete-time control and tracking errors. The evaluation was limited to obstacle-free point-to-point motion and specific initial configurations. Future work is identified as necessary to address broader initial states, obstacle avoidance, and integration with global planning.

More like this