#With the rapid development of artificial intelligence and robotics technology, the importance of functional manipulation in robotics has become increasingly prominent. Traditional benchmark tests can no longer meet the current needs of robots for complex manipulation tasks, calling for the emergence of new manipulation benchmarks (Functional Manipulation Benchmarks). Robot control faces two main challenges: How the robot handles intelligently Complex contact dynamics and how to respond to the diversity of environments and objects. In response to these challenges, robot learning technology is regarded as a key solution. Therefore, the field needs a comprehensive and accessible framework that provides challenging real-world tasks, high-quality data, easily replicable settings, and relevant methods that integrate baseline results. Based on this framework, researchers can conduct experiments on proposed tasks. Discover for in-depth analysis. The research team at the University of California, Berkeley, Robotics Intelligent Laboratory (RAIL) proposed a real-world benchmark as mentioned above, called FMB (Functional Manipulation Benchmark for Generalizable Robotic Learning ). 
- Project homepage: https://functional-manipulation-benchmark.github.io/
- Paper address: https://arxiv.org/abs/2401.08553
- Paper title: FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning
- Co-first author homepage: https://people.eecs.berkeley.edu/~jianlanluo/
- https://charlesxu0124.github.io/
FMB has the following characteristics:
- Innovative design: produced using 3D printing technology Objects in the task are used to test the robot's generalization ability. This method can also be easily reproduced by other researchers.
- Diversified tasks: including single-object and multi-object multi-stage manipulation tasks, truly simulating challenges in daily environments.
- Large Dataset: Through a large number of human demonstrations, the robot is provided with a rich data set.
- Imitation Learning Baseline: Using state-of-the-art machine learning methods, baseline results and modular components are provided for use by other researchers.
The tasks in FMB are roughly Divided into two categories: single-object multi-step manipulation tasks and multi-object multi-step manipulation tasks. These tasks are designed to test the robot's basic skills such as grasping, repositioning and assembly, which are necessary to complete the entire task. The tasks in FMB require the robot to not only complete a single control skill, but also require the robot to combine these skills to complete more complex multi-step tasks. FMB's task design is flexible and changeable. Researchers can choose to focus on a single skill as needed, study the robot's control capabilities in depth, or study complete multi-step tasks. This requires long-term planning on the part of the robot and the ability to recover from failure. More complex multi-step tasks require the robot to make complex real-time decisions, as they involve selecting appropriate objects and reasoning about the sequence of manipulating them.
In the process of robot learning, the role of data cannot be underestimated. In order to enable robots to better understand and master complex tasks, the research team collected a large-scale expert human demonstration data set covering the above tasks, containing more than 20,000 operation trajectories. The research team used four different cameras to record these demonstration data, two of which were mounted on the robot's wrist and two of which provided a global perspective. These cameras capture data such as RGB color image information, depth information, and more that are critical for the robot to learn to solve tasks. In addition, the data set also records force/torque information of the robot’s end effector, which is very important for tasks like assembly that require contact with a large number of objects. Through this rich data, robots can deeply understand every detail of the task and imitate human operating skills more accurately. It is precisely because of the depth and breadth of data that it provides a solid foundation for robot learning. This enables robots to respond to tasks more humanely and dexterously when performing complex tasks.
Imitation Learning Baseline
Architecture diagram of the baseline strategy. Both models based on Transformer and ResNet use a ResNet encoder with shared weights to encode each image view, and then combine it with proprioception Information and selectable objects are combined with corresponding robot skill encoding features to predict 7 degrees of freedom actions. #The experimental part of FMB conducts a series of tests on the performance of imitation learning systems, comparing different learning methods and exploring the impact of different input modes and design decisions. Experiments found that using depth information helps improve the effectiveness of grasping strategies, and force/torque information is very important for assembly tasks. For multi-step tasks, traditional ResNet, Transformer and Diffusion methods have failed, but the hierarchical control method proposed in this paper shows potential.
The experimental results show that the ResNet strategy that incorporates depth information is The performance in the crawling task is consistently better than the strategy using only RGB information. Through the data reduction study, the research team explored the impact of different amounts of training data on the performance of the crawling task. The results show that the performance of the ResNet strategy that incorporates depth information when processing seen objects will improve as the amount of training data increases. Notably, this strategy shows similar performance to seen objects for unseen objects, indicating that the diversity of training objects greatly contributes to the robot's generalization ability.
The importance of force/torque information in assembly tasks It was confirmed. Force/torque information is very important for the strategy adopted by the robot to determine whether the object has contacted the target surface and to effectively conduct actions such as searching.
#However, when the policy is trained on all objects, the robot is not always able to successfully complete the assembly task. This is because the strategy needs to first determine which hole the object should be fitted into and then generate the corresponding actions, which greatly increases the complexity of the task. In order to solve this problem, the research team added an object selection mechanism to the strategy to help the strategy determine the shape of the objects that need to be assembled, thereby focusing on generating correct assembly actions.
FMB’s framework encompasses two complex tasks. These complex tasks require robots to be able to complete multiple steps in a row just like humans. The previous method was to let the robot learn the entire process, but this method was prone to accumulating errors due to errors in a single link, eventually leading to the failure of the entire task. This approach has a success rate of 0/10 in both single and multiple object manipulation tasks. To address the cumulative error problem, the research team adopted a hierarchical control strategy. The hierarchical strategy decomposes the task into several small pieces. Each completed piece is equivalent to passing a decision point. Even if errors occur, they can be quickly corrected to avoid affecting subsequent links. For example, if a robot fails to securely grasp an object during a grasp, it will keep trying until it succeeds. The research team tested two hierarchical approaches. The first provides a valid vector indicating the task type for a single policy, while the second provides for each The control skills are trained separately with different strategies, both using the operator's instructions as the upper-level strategy. In the test, the research team found that both methods performed well. The test results show the effectiveness of the hierarchical approach in handling complex robotic tasks and provide new research directions for future research. As shown in the picture above, the robot can autonomously perform functional control after learning.
Overall, the above experiments demonstrate the technological innovation of the research team in the field of robot learning, and also verify that FMB is a benchmark suitable for developing advanced robot learning methods. The research team looks forward to future research that can further push the boundaries of robot learning based on FMB. The above is the detailed content of Berkeley open source high-quality large-scale robot control benchmark, no longer difficult to face complex autonomous control tasks. For more information, please follow other related articles on the PHP Chinese website!
