A new paper argues that AI agents handling human motion need a better way to keep learning without forgetting what they already know.
Researchers built on a frozen large language model backbone and layered on low-rank adaptation variants — specifically a mixture-of-experts design where an autoencoder-based router picks task-specific experts at inference time, no task label required. They tested the approach on a five-task benchmark derived from the HumanML3D dataset, covering both motion-to-text and text-to-motion directions. Results showed near-zero catastrophic forgetting while keeping generation and captioning quality high. Crucially, hard expert selection — routing inputs to one expert cleanly — beat soft blending across quality metrics, suggesting that keeping experts isolated matters more than letting them share.
Continual learning is a stubborn unsolved problem in AI: models trained sequentially tend to overwrite earlier skills as they absorb new ones. A motion-language agent that loses how to describe a golf swing after learning parkour vocabulary is useless in any real deployment. The isolation finding here echoes similar results in mixture-of-experts research elsewhere, adding weight to the idea that specialization, not averaging, is the right structural bet.
The authors also flag a gap between token-level accuracy and actual generation quality — a reminder that the metrics the field defaults to may not capture what matters, and that leaderboard scores in continual learning research deserve extra scrutiny.