Learning Bimanual Manipulation via Action Chunking and Inter-Arm Coordination with Transformers

Robots that can operate autonomously in a human living environment are necessary to have the ability to handle various tasks flexibly. One crucial element is coordinated bimanual movements that enable functions that are difficult to perform with one hand alone. In recent years, learning-based models that focus on the possibilities of bimanual movements have been proposed. However, the high degree of freedom of the robot makes it challenging to reason about control, and the left and right robot arms need to adjust their actions depending on the situation, making it difficult to realize more dexterous tasks. To address the issue, we focus on coordination and efficiency between both arms, particularly for synchronized actions. Therefore, we propose a novel imitation learning architecture that predicts cooperative actions. We differentiate the architecture for both arms and add an intermediate encoder layer, Inter-Arm Coordinated transformer Encoder (IACE), that facilitates synchronization and temporal alignment to ensure smooth and coordinated actions. To verify the effectiveness of our architectures, we perform distinctive bimanual tasks. The experimental results showed that our model demonstrated a high success rate for comparison and suggested a suitable architecture for the policy learning of bimanual manipulation.

Visual Imitation Learning of Non-Prehensile Manipulation Tasks with Dynamics-Supervised Models

Unlike quasi-static robotic manipulation tasks like pick-and-place, dynamic tasks such as non-prehensile manipulation pose greater challenges, especially for vision-based control. Successful control requires the extraction of features relevant to the target task. In visual imitation learning settings, these features can be learnt by backpropagating the policy loss through the vision backbone. Yet, this approach tends to learn task-specific features with limited generalizability. Alternatively, learning world models can realize more generalizable vision backbones. Utilizing the learnt features, task-specific policies are subsequently trained. Commonly, these models are trained solely to predict the next RGB state from the current state and action taken. But only-RGB prediction might not fully-capture the task-relevant dynamics. In this work, we hypothesize that direct supervision of target dynamic states (Dynamics Mapping) can learn better dynamics-informed world models. Beside the next RGB reconstruction, the world model is also trained to directly predict position, velocity, and acceleration of environment rigid bodies. To verify our hypothesis, we designed a non-prehensile 2D environment tailored to two tasks: "Balance-Reaching" and "Bin-Dropping". When trained on the first task, dynamics mapping enhanced the task performance under different training configurations (Decoupled, Joint, End-to-End) and policy architectures (Feedforward, Recurrent). Notably, its most significant impact was for world model pretraining boosting the success rate from 21% to 85%. Although frozen dynamics-informed world models could generalize well to a task with in-domain dynamics, but poorly to a one with out-of-domain dynamics.