Joint Position Bounds in Resolved-Acceleration Control: A Comparison

Andrea Testa, Luigi Raiano, Marco Laghi, Arash Ajoudani, Enrico Mingo HoffmanHfr 2023 · DOI

TL;DR

This paper tackles the challenge of enforcing joint position, velocity, and acceleration limits in robot control, which is crucial for safe human-robot collaboration. The authors compare three methods (P-Step Ahead Predictor, Control Barrier Functions, and Invariance) and find that a combination of the P-Step Ahead Predictor with the Invariance approach offers the best balance of tracking performance and constraint satisfaction, as demonstrated in a pick-and-place task.

What Problem Does This Paper Address?

When robots work alongside humans, ensuring their movements are safe and predictable is paramount. This involves respecting the robot's physical limitations, such as the maximum range of motion for each joint (joint limits). However, implementing these limits in real-time controllers, especially those that rely on predicting future states (like resolved-acceleration control), is tricky. Simply enforcing limits at each control step isn't enough because the robot's state (position, velocity, acceleration) can change rapidly. If a controller only considers the current state, it might command accelerations that, even if seemingly safe now, will lead to exceeding joint limits in the near future. This can cause damage to the robot or, worse, lead to unsafe interactions with humans. The paper aims to compare different advanced methods for handling these "look-ahead" constraints to find the most effective solution.

Methods

The paper compares three main approaches for enforcing joint limits in a Quadratic Programming (QP)-based resolved-acceleration controller:

  • P-Step Ahead Predictor (PSAP): This method extends the prediction horizon by considering the robot's state up to 'p' steps into the future. It essentially "looks ahead" to anticipate potential limit violations and adjusts the current control commands accordingly. While simple to implement, it requires careful tuning of the 'p' parameter and can lead to conservative behavior.
  • Control Barrier Functions (CBFs): CBFs provide a formal framework for guaranteeing safety by ensuring that the robot's state remains within a "safe set." The paper discusses Exponential-CBFs (E-CBFs) which are extended to handle acceleration constraints. CBFs generally require tuning of parameters and can sometimes lead to chattering in the control output.
  • Invariance (or Viability Control): This approach aims to ensure that if the robot's state is within a feasible region, it will remain within that region for all future times. It guarantees that the robot will not leave a "positive invariance region." This method is attractive because it doesn't require tuning parameters, but it can exhibit chattering at the boundaries.

The authors then propose a hybrid approach: PSAP-aided Invariance, which combines the predictive capabilities of PSAP with the formal guarantees of Invariance.

Datasets

Not discussed in the paper. The evaluation is performed through numerical simulations and a practical experiment on a UR5e robot.

Key Insights and Ideas

  • The "look-ahead" problem: Simply enforcing instantaneous constraints on position, velocity, and acceleration is insufficient for controllers that operate in discrete time steps. The controller needs to anticipate future states to prevent violations.
  • Trade-offs in constraint enforcement: Each method has its strengths and weaknesses. PSAP is simple but requires tuning and can be conservative. CBFs offer formal guarantees but also need tuning. Invariance is tuning-free but can exhibit chattering.
  • Hybrid approaches can be powerful: Combining PSAP with Invariance seems to offer the best of both worlds: the predictive power of PSAP to avoid oscillations and the formal safety guarantees of Invariance.
  • Experimental validation is key: The paper moves beyond theoretical comparisons to demonstrate the effectiveness of the chosen method in a real-world pick-and-place task, highlighting its ability to handle challenging scenarios.

Key Results

  • Numerical Comparison:
    • PSAP: Performance improves with increasing 'p', but higher 'p' values lead to worse tracking. A 'p' value of 100 was selected for comparison.
  • CBF: Parameter 'a' was set to 15 for comparison. CBF showed better tracking than PSAP while maintaining constraints.
  • Invariance: Demonstrated the best tracking performance and constraint satisfaction.
  • Experimental Results (Pick-and-Place Task):
    • The PSAP-aided Invariance approach (with p=2) was implemented on a UR5e robot.
  • The robot successfully completed a pick-and-place task, including navigating to an awkward pick-up location.
  • Crucially, the joint position limits were consistently satisfied throughout the task, with only minor acceleration "chattering" observed when limits were approached.
  • The experiment showed that the chosen method effectively prevented collisions and ensured safe operation.

Limitations and Future Work

  • Limitations:
    • The paper notes that the PSAP approach, while effective, can lead to worse tracking performance compared to other methods.
  • While the PSAP-aided Invariance method shows promise, the paper acknowledges that the basic Invariance approach can exhibit chattering.
  • The experimental setup involved carefully adjusted original joint position limits, suggesting that real-world scenarios might require further fine-tuning or more sophisticated adaptive limit strategies.
  • Future Work:
    • The authors plan to extend the selected method (PSAP-aided Invariance) to handle more complex constraints, such as virtual walls in Cartesian space and collision avoidance.
  • Investigating the application of this approach to systems with a relative degree greater than 1 is also a future direction.

Summary updated

BibTeX
@inproceedings{testa2023joint,
  author = {Andrea Testa and Luigi Raiano and Marco Laghi and Arash Ajoudani and Enrico Mingo Hoffman},
  title = {{Joint Position Bounds in Resolved-Acceleration Control: A Comparison}},
  booktitle = {Hfr},
  year = {2023},
  doi = {10.1007/978-3-031-55000-3\_9},
}
Figure