Conceptos Básicos
Deep Predictive Model with Parametric Bias (DPMPB) is a learning-based approach that can handle complex modeling difficulties and temporal changes in robot dynamics, enabling adaptive control of robots in real-world environments.
Resumen
The paper introduces Deep Predictive Model with Parametric Bias (DPMPB) as a framework to address two key challenges in robot control: modeling difficulties and temporal model changes.
Modeling Difficulties:
- Robots with complex, flexible, or redundant bodies (e.g., musculoskeletal, low-rigidity, flexible hands)
- Handling flexible tools and deformable objects
- Interacting with undefined or changing environments (e.g., floor material, obstacles)
Temporal Model Changes:
- Changes in robot body state due to aging, configuration, or what the robot wears
- Changes in grasped objects, handled tools, or manipulated materials
- Changes in the environment, such as floor, obstacles, or friction
DPMPB uses a neural network to model the correlation between sensors and actuators, allowing it to handle complex relationships. It also applies parametric bias (PB) to implicitly embed information about the current state of the robot, objects, and environment. This enables DPMPB to adapt to temporal changes by online updating of the PB.
The paper classifies predictive models, modeling difficulties, and temporal model changes, and demonstrates the effectiveness of DPMPB through various robot experiments, including:
- Grasping object recognition and contact control of a flexible hand
- Visual feedback control of a low-rigidity robot considering body changes
- Stable control of a wheeled robot considering floor changes
- Balance control of a humanoid robot considering shoe changes
- Dynamic cloth manipulation by a musculoskeletal humanoid
The experiments show that DPMPB can handle the modeling difficulties and temporal changes by simply adjusting the network input/output and a few parameters, without the need to redesign the entire system.
Estadísticas
"When the relationship among the body, tools, target objects, and environment is complex, it is difficult to model using classical methods."
"When the relationship among the body, tools, target objects, and environment changes with time, it is necessary to deal with the temporal changes in the model."
"Robots need a learning system resembling human adaptive intelligence that allows them to cope with these problems in the real world."
Citas
"By introducing parametric bias, changes in the body and environment can be embedded in small dimensional variables, which can be updated online to adapt to the current body, tool, and environment quickly without destroying the dynamics of the entire network."
"DPMPB can realize various tasks by combining its structure (STM or CTM), the handled modeling difficulties among behavior, body, object/tool, and environment, and the handled temporal model changes in behavior, body, object/tool, and environment."