Abstract
Autonomous robots require online trajectory planning capability to operate in the real world. Efficient offline trajectory planning methods already exist, but are computationally demanding, preventing their use online. In this paper, we present a novel algorithm called Guided Trajectory Learning (GTL) that learns a function approximation of solutions computed through trajectory optimization while ensuring accurate and reliable predictions. This function approximation is then used online to generate trajectories. This algorithm is designed to be easy to implement and practical since it does not require massive computing power. It is readily applicable to any robotics systems and effortless to set up on real hardware since robust control strategies are usually already available. We demonstrate the computational performance of our algorithm on flat-foot walking with the self-balanced exoskeleton Atalante.
Introduction: The Feasibility Gap in Robotics
In robotics, there is a classic tension between optimality and speed. To make a complex robot like a bipedal exoskeleton walk safely, engineers use trajectory optimization—mathematical solvers that find the best path while respecting physics and hardware limits. However, these solvers are computationally heavy and often fail to run in real-time (online).
A common workaround is to pre-calculate thousands of trajectories offline and train a neural network to mimic them via regression. However, standard neural networks often produce slightly infeasible paths. For a human in an exoskeleton, a 1-degree error in an ankle joint can lead to a potential fall. This paper addresses the "feasibility gap": ensuring a fast, learned model produces trajectories that the robot can physically execute.
Methods: Guided Trajectory Learning (GTL)
The authors propose Guided Trajectory Learning (GTL), which treats the problem as a consensus optimization using the Alternating Direction Method of Multipliers (ADMM). This iterative process involves a "negotiation" between the trajectory optimizer and the neural network. If the neural network struggles to learn a specific movement, the optimizer is guided to find a different, slightly modified solution for that task that is easier for the network to represent.
Function Approximation Architecture
The system utilizes a Deconvolutional Neural Network specifically designed to generate multi-dimensional temporal sequences (such as the 12-joint angles of a walking gait) from low-dimensional task descriptions.
For practical deployment, the authors highlight a simplified version called GTL-0, which focuses on the penalized reconstruction cost, making it easier to scale across large computing clusters.
Experimental Setup and Datasets
The study uses a custom-generated dataset tailored to the Atalante exoskeleton, encompassing a 12-dimensional task space including patient morphology (height, weight), exoskeleton mechanical settings (thigh/shank lengths), and gait parameters (step length, duration). The dataset consists of 70,000 uniformly sampled tasks, validated through simulation and real-world clinical trials with six human subjects of varying body types.
Key Results and Performance
The GTL-0 algorithm demonstrated significant improvements in both accuracy and stability compared to standard regression techniques. A critical benchmark was the "stability criterion," defined as a joint angle error of less than 0.01 radians.
- Accuracy: Standard regression failed to meet the stability criterion in over 50% of test cases. GTL-0 reduced this failure rate to just 10.5% after only two iterations.
- Real-World Stability: In clinical trials, trajectories generated by standard regression were often unstable on real hardware. In contrast, GTL-0 allowed the Atalante exoskeleton to successfully control average velocity across different subjects, with measured velocities closely matching desired inputs.
- Computational Efficiency: Once trained, the neural network generates trajectories at a fraction of the cost of the original optimization, enabling true online planning.
Discussion and Future Work
A profound insight from this research is that we should not just try to make models smarter; we should make the "expert" (the optimizer) more flexible. By forcing the optimizer to output "learnable" trajectories, the resulting neural network becomes significantly more reliable. Furthermore, GTL smooths out discontinuities in standard optimization, ensuring the robot's behavior changes predictably as the environment changes.
While GTL handles the representation gap well, future work will focus on explicitly modeling the "reality gap" between simulation and real-world hardware. The authors suggest applying GTL to more dynamic scenarios, such as perturbation recovery and path replanning in cluttered environments, potentially using adaptive neural network architectures.
Summary updated