New Stepwise Strategy Boosts Accuracy in Flexible-Joint Robot Calibration
In the rapidly evolving world of robotics, precision isn’t just a goal—it’s a necessity. Whether assembling microelectronics or performing delicate surgical procedures, robots must operate with unwavering accuracy. Yet achieving that level of fidelity hinges on one often-overlooked foundation: dynamic parameter identification. Without correctly calibrated models of inertia, friction, gravity, and joint flexibility, even the most advanced control algorithms falter. Now, a team of researchers from Nanjing University of Aeronautics and Astronautics has unveiled a novel stepwise parameter identification strategy specifically tailored for robots equipped with flexible joints—a breakthrough that could reshape how engineers approach robotic calibration in real-world applications.
Flexible-joint robots, which use compliant elements like harmonic drives to absorb shocks and improve safety, present unique challenges compared to their rigid counterparts. The elasticity introduced by these components creates complex dynamic interactions between motors, gears, and links. Traditional identification methods—often designed for stiff mechanical systems—struggle to disentangle these coupled effects, leading to poor model fidelity and suboptimal performance. What’s more, many existing approaches treat parameter identification as a monolithic problem, attempting to solve for all unknowns simultaneously. This “all-at-once” tactic frequently results in ill-conditioned mathematical systems where small measurement errors balloon into massive parameter inaccuracies.
The new method, detailed in a recent study published in Mechanical & Electrical Engineering Magazine, sidesteps these pitfalls by breaking the identification process into a carefully orchestrated sequence. Rather than brute-forcing a solution, the researchers propose a two-pronged analytical framework grounded in linear algebra: first, they assess whether subsets of parameters can be uniquely determined (a question of matrix rank); second, they evaluate how sensitive those solutions are to noise and disturbances (a matter of condition number). Only when both criteria are satisfied does a parameter group proceed to identification.
“This isn’t just about better math—it’s about smarter experimentation,” explains Feixiang Chen, lead author of the study and a graduate researcher at Nanjing University of Aeronautics and Astronautics. “By strategically ordering which parameters we identify first, and designing motion trajectories that maximize numerical stability, we dramatically reduce error propagation. It’s like solving a puzzle piece by piece instead of shaking the whole box and hoping everything falls into place.”
At the heart of the strategy lies a refined dynamic model that explicitly accounts for gravity compensation using joint torque sensors—an increasingly common feature in modern collaborative robots. Unlike prior work that relied on external force/torque sensors mounted at the end-effector, this approach leverages internal torque feedback to isolate gravitational effects directly at each joint. This not only simplifies hardware requirements but also enhances robustness in scenarios where external sensing is impractical or unreliable.
The team validated their method on a two-degree-of-freedom flexible-joint robot, a representative platform that captures the core complexities of multi-link compliant systems. They began by designing smooth, periodic excitation trajectories—essentially choreographed motions—that would sufficiently “excite” the robot’s dynamics without exceeding mechanical limits. Using an eighth-order polynomial form wrapped into a repeating cycle, they ensured zero velocity and acceleration at trajectory endpoints, enabling safe, repeatable experiments.
Crucially, the researchers didn’t settle for any arbitrary trajectory. Instead, they employed a genetic algorithm to optimize the coefficients of these motion profiles, minimizing the condition number of the resulting observation matrices. In practical terms, this means the chosen movements were specifically engineered to yield the most numerically stable parameter estimates possible. The difference is stark: poorly conditioned systems might amplify sensor noise by factors of thousands, whereas the optimized trajectories kept amplification below a factor of ten—a threshold widely regarded as acceptable in engineering practice.
The stepwise workflow unfolds as follows. First, friction parameters—both viscous and Coulomb—are identified independently by commanding slow, quasi-static joint motions where inertial and Coriolis effects are negligible. Once these nonlinear friction terms are quantified, they’re subtracted from subsequent datasets, effectively “linearizing” the remaining dynamics. Next, the robot executes single-joint motions to identify partial inertial and gravitational parameters. Only after these foundational elements are secured does the system move to coordinated two-joint motions, where coupling effects become pronounced and full model completion becomes feasible.
One particularly clever insight involves the treatment of motor-side versus link-side dynamics. Because harmonic drives typically feature high reduction ratios (101:1 in the test setup), rotor speeds are drastically lower than link speeds. This allows certain cross-coupling terms in the kinetic energy equations to be safely neglected—a simplification that preserves model accuracy while drastically reducing computational complexity. Moreover, by assuming rotor centers of mass align with rotation axes, gravitational potential terms simplify further, streamlining the identification of mass-related parameters.
The experimental results speak volumes. All optimized observation matrices achieved condition numbers well below 6—far beneath the critical threshold of 10. More importantly, the sequential approach prevented error accumulation that plagues simultaneous identification schemes. When the team attempted to merge motor and link parameter estimation into a single step, the condition number skyrocketed due to near-singularities in the combined matrix—a clear warning sign that such fusion was numerically hazardous. By respecting the inherent structure of the dynamics, the stepwise method avoided this trap entirely.
Beyond its immediate application to flexible-joint arms, the framework exhibits remarkable generality. The authors emphasize that their rank-and-condition-number methodology applies to any linear-in-parameters dynamic system—not just robots, but potentially aerospace mechanisms, prosthetic limbs, or even biomechanical models. As long as the governing equations can be cast into a linear regression form (a surprisingly broad class of problems), the same principles of sequential decoupling and trajectory optimization hold.
Industry experts see immediate implications. “Calibration time is a major bottleneck in deploying industrial robots,” notes a senior robotics engineer at a leading automation firm who reviewed the work. “If you can cut identification from hours to minutes while improving accuracy, that’s a game-changer for both manufacturers and integrators. This paper offers a principled path to do exactly that.”
Indeed, the economic stakes are high. In automotive assembly lines, for instance, a miscalibrated robot might apply excessive force during part insertion, damaging components worth thousands of dollars. In medical robotics, inaccurate dynamic models could compromise patient safety during assisted procedures. Even in consumer-facing applications like warehouse logistics, poor calibration leads to jerky motions, reduced throughput, and premature wear.
What sets this work apart is its balance of theoretical rigor and practical pragmatism. While rooted in solid mathematical foundations—matrix theory, perturbation analysis, Lagrangian mechanics—the proposed strategy never loses sight of real-world constraints. Trajectories respect joint limits and actuator capabilities. Sensor noise is acknowledged as inevitable, not ignored. And the entire workflow fits within standard experimental protocols used in research labs and production facilities alike.
Looking ahead, the researchers suggest several natural extensions. Integrating online adaptation could allow robots to continuously refine their models during operation, compensating for wear, temperature drift, or payload changes. Combining the stepwise approach with machine learning might further accelerate convergence, using neural networks to predict promising initial parameter guesses. And scaling the method to six- or seven-degree-of-freedom arms—while computationally more demanding—appears entirely feasible given the modular nature of the identification sequence.
For now, the study stands as a compelling blueprint for next-generation robot calibration. It reminds us that in an era obsessed with artificial intelligence and deep learning, sometimes the most impactful advances come not from adding complexity, but from thoughtfully structuring what we already know. By returning to first principles and asking not just what to identify but how and when to identify it, Chen and his colleagues have delivered a strategy that is as elegant as it is effective.
As flexible-joint robots continue their march from research labs into factories, hospitals, and homes, reliable parameter identification will only grow in importance. Thanks to this work, engineers now have a clearer, more robust roadmap to ensure these machines move not just with power, but with precision.
Feixiang Chen, Chenggang Li, Yadong Chu, Jian Wang, Yunshuang Yue
College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
Mechanical & Electrical Engineering Magazine
DOI: 10.19344/j.cnki.issn1671-5276.2021.06.042