Robot control using trajectories
By employing a 2D trajectory sketch to guide robot actions, the system overcomes the challenge of generalizing to unseen tasks and motions, achieving improved controllability and performance in diverse scenarios.
Patent Information
- Application Number
- PCT/EP2024/074229
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-09-29
- Filing Date
- 2024-08-29
- Publication Date
- 2025-04-03
AI Technical Summary
Existing approaches to robot control using neural networks face challenges in generalizing to tasks not seen in training data, particularly when trained using imitation learning from expert demonstrations. Additionally, previous methods suffer from a lack of specificity and difficulty in generalizing to new motions.
The system uses a 2D trajectory sketch to guide the robot's actions, which is deliberately underspecified and provides approximate guidance. This approach allows the neural network to perform new tasks and generalize to novel motions, even when trained only using imitation learning.
The described techniques enable the system to generalize well to diverse new tasks, motions, and objects, improving controllability and performance beyond the training data. The use of a trajectory sketch facilitates easy training and utilization of training data, allowing the system to adapt to new situations.