The intersection of robotics and artificial intelligence led to a profound paradigm shift in Robot Learning. Robots have the capacity to replicate human actions and also dynamically adapt, innovate, and excel across a spectrum of tasks. However, the heterogeneity in the deployment of robot platforms and software frameworks poses considerable challenges in terms of systematic testing and comparative analyses. Additionally, the data scarcity of especially force controlled robot manipulation is still restraining the development of advanced foundation models. A reference platform with default software stack can help to increase comparability, reducing development time and collect a large amount of tactile robot manipulation data. To address on this problem, we developed a Parallel and Distributed Robot AI (PD.RAI) framework, comprising a scalable ensemble of Robot Learning Units (RLUs), a global database, and the Robot Cluster Intelligence (RoCI). Each RLU is endowed with robot arms, cameras, and local computational units to autonomously engage in planning, control, and local machine learning of tactile manipulation skills. The RoCI system oversees the learning process and schedules the RLUs tasks. To show the functionality of the system, two black-box optimization algorithms are compared within the robot skill learning domain. An experiment with 24 different optimization tasks is conducted in parallel. The algorithms are incorporated into the same existing default modules acting as a reference environment. This allows for a realistic comparison without sacrificing diversity of possible configurations and testing environments.

Robotic assembly tasks are open challenges due to the long task horizon and complex part relations. Behavior trees (BTs) are increasingly used in robot task planning for their modularity and flexibility, but creating them can be effortintensive. Large language models (LLMs) have recently been applied in robotic task planning for generating action sequences, but their ability to generate BTs has not been investigated. To this end, We propose LLM as BT-planner, a novel framework to leverage LLMs for BT generation in robotic assembly task planning. Four in-context learning methods are introduced to utilize the generative and natural language processing capabilities of LLMs to produce task plans in BT format, reducing manual effort and ensuring robustness and comprehensibility. We also evaluate the performance of fine-tuned fewer-parameter LLMs on the same tasks. Experiments in simulated and real-world settings show that our framework enhances LLMs' performance in BT generation, improving success rates in BT generation through in-context learning and supervised fine-tuning.

Sequential manipulation planning has been a critical imperative to achieve a higher level of autonomy in robotics. Classical approaches to address task planning problems are based on symbolic formalisms, such as Planning Domain Definition Language (PDDL) [1], and search for state transition plans to reach task goals. In practice, such task plans are often programmed as Finite State Machines (FSMs), which incorporate expert knowledge specifying control and execution details. Due to its limitation of scalability [2], Behavior trees (BTs), which represent policies in a state-less, hierarchical tree structure, have gained increasing popularity for complex task planning. Its advantages of modularity, reusability and reactivity, make it a more desired formalism for long-horizon manipulation tasks.