Structural understanding of complex visual objects is an important unsolved component of artificial intelligence. To study this, we develop a new technique for the recently proposed Break-and-Make problem in LTRON where an agent must learn to build a previously unseen LEGO assembly using a single interactive session to gather information about its components and their structure. We attack this problem by building an agent that we call \textbf{\ours} that is able to make its own visual instruction book. By disassembling an unseen assembly and periodically saving images of it, the agent is able to create a set of instructions so that it has the information necessary to rebuild it. These instructions form an explicit memory that allows the model to reason about the assembly process one step at a time, avoiding the need for long-term implicit memory. This in turn allows us to train on much larger LEGO assemblies than has been possible in the past. To demonstrate the power of this model, we release a new dataset of procedurally built LEGO vehicles that contain an average of 31 bricks each and require over one hundred steps to disassemble and reassemble. We train these models using online imitation learning which allows the model to learn from its own mistakes. Finally, we also provide some small improvements to LTRON and the Break-and-Make problem that simplify the learning environment and improve usability.

When modeling complex robot systems such as branched robots, whose kinematic structures are a tree, current techniques often require modeling the whole structure from scratch, even when partial models for the branches are available. This paper proposes a systematic modular procedure for the dynamic modeling of branched robots comprising several subsystems, each composed of an arbitrary number of rigid bodies, providing the final dynamic model by reusing previous models of each branch. Unlike previous approaches, the proposed strategy is applicable even if some subsystems are regarded as black boxes, requiring only twists and wrenches at the connection points between them. To help in the model composition, we also propose a weighted directed graph representation where the weights encode the propagation of twists and wrenches between the subsystems. A simple linear operation on the graph interconnection matrix provides the dynamics of the whole system. Numerical results using a 24-DoF fixed-base branched robot composed of eight subsystems show that the proposed formalism is as accurate as a state-of-the-art library for robotic dynamic modeling. Additional results using a 30-DoF holonomic branched mobile manipulator composed of three subsystems demonstrate the fidelity of our model to a modern robotics simulator and its capability of dealing with black box subsystems. To further illustrate how the derived dynamic model can be used in closed-loop control, we also present a simple formulation of a model-based wrench-driven pose control for branched robots.

Developments in touch-sensitive textiles have enabled many novel interactive techniques and applications. Our digitally-knitted capacitive active sensors can be manufactured at scale with little human intervention. Their sensitive areas are created from a single conductive yarn, and they require only few connections to external hardware. This technique increases their robustness and usability, while shifting the complexity of enabling interactivity from the hardware to computational models. This work advances the capabilities of such sensors by creating the foundation for an interactive gesture recognition system. It uses a novel sensor design, and a neural network-based recognition model to classify 12 relatively complex, single touch point gesture classes with 89.8% accuracy, unfolding many possibilities for future applications. We also demonstrate the system's applicability and robustness to real-world conditions through its performance while being worn and the impact of washing and drying on the sensor's resistance.

Visual understanding of geometric structures with complex spatial relationships is a fundamental component of human intelligence. As children, we learn how to reason about structure not only from observation, but also by interacting with the world around us -- by taking things apart and putting them back together again. The ability to reason about structure and compositionality allows us to not only build things, but also understand and reverse-engineer complex systems. In order to advance research in interactive reasoning for part-based geometric understanding, we propose a challenging new assembly problem using LEGO bricks that we call Break and Make. In this problem an agent is given a LEGO model and attempts to understand its structure by interactively inspecting and disassembling it. After this inspection period, the agent must then prove its understanding by rebuilding the model from scratch using low-level action primitives. In order to facilitate research on this problem we have built LTRON, a fully interactive 3D simulator that allows learning agents to assemble, disassemble and manipulate LEGO models. We pair this simulator with a new dataset of fan-made LEGO creations that have been uploaded to the internet in order to provide complex scenes containing over a thousand unique brick shapes. We take a first step towards solving this problem using sequence-to-sequence models that provide guidance for how to make progress on this challenging problem. Our simulator and data are available at github.com/aaronwalsman/ltron. Additional training code and PyTorch examples are available at github.com/aaronwalsman/ltron-torch-eccv22.

Part assembly is a typical but challenging task in robotics, where robots assemble a set of individual parts into a complete shape. In this paper, we develop a robotic assembly simulation environment for furniture assembly. We formulate the part assembly task as a concrete reinforcement learning problem and propose a pipeline for robots to learn to assemble a diverse set of chairs. Experiments show that when testing with unseen chairs, our approach achieves a success rate of 74.5% under the object-centric setting and 50.0% under the full setting. We adopt an RRT-Connect algorithm as the baseline, which only achieves a success rate of 18.8% after a significantly longer computation time. Supplemental materials and videos are available on our project webpage.

It's Friday September 14, 2046, and you are at the airport getting an alert from your car that your 10,000 mile service is due, but something peculiar comes with your alert. It's a question: your car asks you if you wish to have the service taken care of within the next 10 days, or after. Your car's smart system contacts the car dealer's smart system and arranges everything, but first it checks your calendar and figures out if you are out of town on any particular day. It discovers you will be traveling for the day on Monday the 24th, so it books your car service to the airport from the service department, your airline and hotel reservations, and your dinner reservations with your client. It also updates your social profiles for you while sending you relevant information about where you are going, whom you know there, whom you have not been in touch with for a while that may be in the area, and a plethora of other intelligence useful for you.