Vision-Language-Action Policies (such as Gr00T, Pi0, OpenVLA) are designed so that an adaptation stage is needed before reliable deployment (reliable meaning with very high success rate) of the policy model on a specific combination of application environment and robotic task.
This adaptation stage, often called fine-tuning or post-training, caters to the narrow set of tasks, done in the exact deployment environment, atop a specific embodiment (such as single arm robots, or bi manual robots, or humanoids among others).
The recently released World-Action-Models (WAMs), while having improved generalization to novel tasks and environments, also need adaptation stage to reliably execute a given task on a given environment.
The key requirement of robotic adaptation is high quality adaptation data, typically human teleoperation data of the robot performing the required task in the given environment. Typically 50-200 examples of human teleoperation are required, depending on the base policy being finetuned, as well as the desired performance levels and the complexity of the task.
Thus, given any change in the environment or task where we want robot to perform reliably, there will be a need to collect such teleoperation data again and again.
Thus, robotic adaptation, which is the critical last mile of reliable robotic policy deployment - suffers from the data collection bottleneck - which needs to be eased.
Thus, an important research question needing solutions is -
What are effective approaches to ease human-teleoperation data collection effort required for robotic policy adaptation?