New Robot Learning Method Helps Robots Master New Tasks Without Forgetting Old Ones
This could lead to robots that continuously learn and adapt in your home or workplace.
HIDRA is a structured routing framework for replay-free, single-pass lifelong imitation learning, where past data cannot be stored or revisited. It uses a hierarchical dual-attention mechanism to decouple instruction-level expert retrieval from context-dependent refinement, plus key-level geometric regularization that enforces alignment within tasks and separation across tasks, enabling reliable expert selection and an expandable set of experts for continual adaptation. Evaluated on several LIBERO benchmark suites, including variants with heterogeneous task sequences and paraphrased language instructions, HIDRA consistently outperforms both replay-free and replay-based baselines, improving AUC and forward transfer while maintaining low forgetting, with stronger gains under high interference and language variation.
- HIDRA enables robots to learn new tasks without forgetting old ones, even when they can't store past data.
- It uses a dual-attention system to keep expert selection stable and a geometric regularization to prevent interference.
- In tests, HIDRA outperformed other methods, especially with varied tasks and language instructions.
Why It Matters
Robots that learn continuously could soon handle diverse tasks in homes and workplaces, saving time and money.