Connect with top gaming leaders in Los Angeles at GamesBeat Summit 2023 this May 22-23. Register here.
In a pair of papers published on the preprint server Arxiv.org this week, Google and University of California, Berkeley researchers describe new AI and machine learning techniques that enable robots to adapt to never-before-seen tasks and grasp occluded objects. The first study details X-Ray, an algorithm that when deployed on a robot can search through heaps of objects to grasp a target object, while the second lays out a policy adaptation technique that “teaches” robots skills without requiring from-scratch model training.
Robot grasping is a surprisingly difficult challenge. For example, robots struggle to perform what’s called “mechanical search,” which is when they have to identify and pick up an object from within a pile of other objects. Most robots aren’t especially adaptable, and there’s a lack of sufficiently capable AI models for guiding robot hands in mechanical search.
X-Ray and the policy adaptation step could form the foundation of a product-packaging system that spots, picks up, and drops a range of objects without human oversight.
The coauthors of the study about X-Ray note that mechanical search — finding objects in a heap of objects — remains challenging due to a lack of appropriate models. X-Ray tackles the problem with a combination of occlusion inference and hypothesis predictions, which it uses to estimate an occupancy distribution for the bounding box (coordinates for a rectangular border around an object) most similar to an object while accounting for various translations and rotations.
Join us in San Francisco on July 11-12, where top executives will share how they have integrated and optimized AI investments for success and avoided common pitfalls.
X-Ray assumes that there’s at least one target object fully or partially occluded by unknown objects in a heap, and that a maximum of one object is grasped per timestep. Taking RGB images and target objects as inputs, it predicts the occupancy distribution and segmentation masks for the scene and computes several potential grasping actions, executing the one with the highest probability of succeeding.
To train and validate X-Ray, the researchers produced a corpus of 10,000 augmented depth images labeled with object occupancy distributions for a rectangular box target object. Sampling from an open source data set of 1,296 3D CAD models on Thingiverse, they selected 10 box targets of various dimensions with equal volume but small thickness, so that they were more likely to be occluded. This netted them a total of 100,000 images.
About 8,000 of those 10,000 images were reserved for training, and the rest were set aside for testing. One thousand additional images containing simulated objects — a lid, a domino, and a flute — were used to evaluate X-Ray’s generalization to unseen shapes, objects, aspect ratios, and scales.
In physical experiments involving a real-world ABB YuMi robot with a suction cup and a parallel jaw gripper, the researchers tasked X-Ray with filling a bin with objects and then dumping the bin on top of the target object. In heaps initially containing 25 objects, the system extracted the target object in a median of 5 actions over 20 trials with a 100% success rate.
The coauthors leave to future work increasing X-Ray’s training efficiency and analyzing the effect of data set size and the number of translations and rotations used to generate training distributions. They also plan to explore reinforcement learning policies based on the reward of target object visibility.
In the more recent of the two papers, the coauthors sought to develop a system that continuously adapts to new real-world environments, objects, and conditions. That’s in contrast to most robots, which are trained once and deployed without much in the way of adaptation capabilities.
The researchers pretrained a machine learning model to grasp a range of objects on a corpus of 608,000 grasp attempts, which they then tasked with grasping objects using a gripper moved 10 centimeters to the right of where it started. After the system practiced gripping for a while (over the course of 800 attempts) and logged those attempts into a new data set — a target data set — the new attempts were mixed in 50% of the time with the original data set to fine-tune the model.
These steps — pretraining, attempting a new task, and fine-tuning — were repeated for five different scenarios. In one, harsh lighting impeded the robot’s cameras; in another, a checkerboard-patterned background made it difficult for the model to identify objects. Lastly, the experimenters had the robot grasp transparent bottles not seen during training (transparent objects are notoriously hard for robots to grasp because they sometimes confuse depth sensors) and pick up objects sitting on a highly reflective sheet metal surface.
The researchers report that in experiments, the model successfully grasped objects 63% of the time in harsh lighting, 74% of the time with transparent bottles, 86% of the time with a checkerboard backing, 88% of the time with an extended gripper, and 91% of the time with an offset gripper. Moreover, they say that it only took 1 to 4 hours of practice for the robot to adapt to new situations (compared with roughly 6,000 hours learning how to grasp) and that performance didn’t degrade the more the model adapted.
In the future, the team plans to investigate whether the process can be made automatic.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.