Continual Online Personalization of Exoskeleton Control via Manifold-Aware Experience Replay

Technical Summary: Continual Online Personalization of Exoskeleton Control via Manifold-Aware Experience Replay

Problem Statement
Personalizing exoskeleton control for clinical users, particularly those with gait disabilities like hemiplegia, remains a critical challenge. While Online Adaptation (OA) offers a solution by adjusting controllers in real-time to user variability and device fit, it faces the issue of catastrophic forgetting. In continual learning scenarios, adapting to a current task often causes the model to overwrite knowledge of previously learned locomotor contexts. Existing experience replay methods in wearable robotics have primarily been applied to prosthetics and rely on explicit task labels to organize replay buffers. However, in real-world exoskeleton deployment, task labels are often unavailable, and sensing modalities are sparser than in prosthetic applications. Furthermore, current frameworks struggle to maintain stability across diverse, multimodal locomotor tasks (e.g., varying speeds and inclines) without explicit context labeling.

Methodology
The authors propose a manifold-aware experience replay framework designed to personalize a hip exoskeleton gait phase estimator for users with emulated hemiplegic gait. The system operates on a robotic hip exoskeleton equipped with bilateral hip encoders and an edge controller (NVIDIA Jetson Orin Nano) running a parallelized software architecture.

1. Gait Phase Estimation: The core controller is a Temporal Convolutional Network (TCN) that estimates gait phase from bilateral hip encoder signals. The network is pre-trained on able-bodied data and serves as the seed for online personalization.
2. Online Adaptation (OA) Strategy: The system updates the model every third heel strike. To preserve general locomotor features while adapting to user-specific traits, the authors freeze the high-level TCN feature-extracting blocks and only fine-tune the final linear layer that maps latent space to gait phase.
3. Manifold-Aware Replay: To prevent catastrophic forgetting without relying on explicit task labels, the framework encodes a gait manifold.
   • Latent Space Construction: Principal Component Analysis (PCA) is applied to kinematic data, retaining the first two principal components (accounting for 77% of variance) and the gait cycle duration as a third axis. This creates a 3D latent space where different tasks (speeds and inclines) form distinct clusters.
   • Binning and Selection: Incoming data is projected into this latent space and discretized into bins based on grid intervals. New bins are created for unseen manifolds, while existing bins are updated with recent data.
   • Replay Mechanism: During adaptation, the system evaluates the Root Mean Square Error (RMSE) of all bins. It selects the top four bins with the lowest accuracy (within a specific RMSE range of 2.5%–20%) to combine with current data for backpropagation. This ensures the model rehearses previously learned tasks that are currently performing poorly, without needing to know the specific task label.

Key Contributions

• Label-Free Replay: The framework eliminates the need for explicit task labeling during replay buffer selection by utilizing a gait manifold derived from kinematic data.
• Multimodal Continual Learning: It successfully demonstrates OA across diverse locomotor tasks (varying speeds and inclines) for a clinical population, a scenario previously unaddressed in exoskeleton studies.
• Architectural Efficiency: By freezing the feature extractor and only updating the final linear layer, the system conserves computational resources and reduces update latency while preserving general human locomotion features.
• Real-Time Performance: The system achieves an average update latency of approximately 1.5 seconds, enabling prompt network updates responsive to user needs.

Results
The framework was evaluated on nine able-bodied subjects emulating hemiplegic gait (using knee braces and metronome cues to induce asymmetry) across six locomotor tasks involving different speeds and inclines. The experiment included a personalization phase, a forgetting phase (exposure to a single task), and a testing phase.

• Torque Tracking: The manifold-aware replay framework achieved a 40% improvement in torque tracking accuracy (RMSE) compared to a baseline framework without replay. Specific improvements ranged from 21% to 56% depending on the task transition pair.
• Gait Phase Estimation: The proposed method achieved a 60% improvement in gait phase tracking accuracy compared to the baseline.
• Forgetting Mitigation: The baseline framework exhibited significant catastrophic forgetting during task transitions, particularly when speed changed. The proposed method maintained stable performance across all task transitions.
• Latency: The update latency plateaued at roughly 1.5 seconds with a maximum of four replay bins, ensuring the system remained responsive.

Significance and Claims
The paper claims that this manifold-aware replay framework enables exoskeletons to deliver personalized assistance in real-time across diverse locomotor contexts encountered during daily ambulation. By addressing the dual challenges of catastrophic forgetting and the lack of explicit task labels, the method allows for stable, continual adaptation to user-specific gait patterns. The authors emphasize that this approach is particularly vital for clinical populations where inter-subject variability is high and access to gait labs for extensive tuning is limited. The study concludes that leveraging the gait manifold allows the controller to maintain stability across learned tasks while adapting to new ones, a crucial step toward robust, real-world exoskeleton deployment.