We introduce SPOT, an object-centric imitation learning framework. The key idea is to capture each task by an object-centric representation, specifically the SE(3) object pose trajectory relative to the target. This approach decouples embodiment actions from sensory inputs, facilitating learning from various demonstration types, including both action-based and action-less human hand demonstrations, as well as crossembodiment generalization. Additionally, object pose trajectories inherently capture planning constraints from demonstrations without the need for manually-crafted rules. To guide the robot in executing the task, the object trajectory is used to condition a diffusion policy. We systematically evaluate our method on simulation and real-world tasks. In real-world evaluation, using only eight demonstrations shot on an iPhone, our approach completed all tasks while fully complying with task constraints. Project page: https://nvlabs.github.io/object_centric_diffusion

We present Asymmetric Critic Guided Distillation, ACGD, a framework for learning multi-task dexterous manipulation policies that can manipulate articulated objects using images as input. ACGD is a scalable student-teacher distillation approach that utilizes behavior cloning to distill multiple expert policies into a single vision-based, multi-task student policy for dexterous manipulation. The expert policies are trained with traditional RL techniques with access to privileged state information of both the robot and the manipulated object, while the distilled student policy operates under realistic sensory constraints, specifically using only camera images and robot proprioception. During distillation, we use an expert-critic that provides action labels and value estimates to refine the student’s action sampling through a dual IL/RL objective. In the multi-task setting, we achieve this through an aggregate critic for different single-task experts. Our approach exhibits strong performance compared to a number of state-of-the-art imitation learning (IL) and reinforcement learning (RL) baselines. We evaluate across a variety of multi-task dexterous manipulation benchmarks including bimanual manipulation, single-hand object articulation tasks, and a tendon-actuated hand and achieves state-of-the-art performance with 10-15% improvement over the baseline algorithms. Visit our website for more details.

For both humans and robots, the sense of touch, known as tactile sensing, is critical for performing contact-rich manipulation tasks. Three key challenges in robotic tactile sensing are interpreting sensor signals, generating sensor signals in novel scenarios, and learning sensor-based policies. For visuotactile sensors, interpretation has been facilitated by their close relationship with vision sensors (e.g., RGB cameras). However, generation is still difficult, as visuotactile sensors typically involve contact, deformation, illumination, and imaging, all of which are expensive to simulate; in turn, policy learning has been challenging, as simulation cannot be leveraged for large-scale data collection. We presentTacSL(taxel), a library for GPU-based visuotactile sensor simulation and learning.TacSLcan be used to simulate visuotactile images and extract contact-force distributions over$200\times$faster than the prior state-of-the-art, all within the widely used Isaac simulator. Furthermore,TacSLprovides a learning toolkit containing multiple sensor models, contact-intensive training environments, and online/offline algorithms that can facilitate policy learning for sim-to-real applications. On the algorithmic side, we introduce a novel online reinforcement-learning algorithm called asymmetric actor-critic distillation, designed to effectively and efficiently learn tactile-based policies in simulation that can transfer to the real world. Finally, we demonstrate the utility of our library and algorithms by evaluating the benefits of distillation and multimodal sensing for contact-rich manipulation tasks, and most critically, performing sim-to-real transfer.

Model-Free Reinforcement Learning (MFRL), leveraging the policy gradient theorem, has demonstrated considerable success in continuous control tasks. However, these approaches are plagued by high gradient variance due to zeroth-order gradient estimation, resulting in suboptimal policies. Conversely, First-Order Model-Based Reinforcement Learning (FO-MBRL) methods employing differentiable simulation provide gradients with reduced variance but are susceptible to sampling error in scenarios involving stiff dynamics, such as physical contact. This paper investigates the source of this error and introduces Adaptive Horizon Actor-Critic (AHAC), an FO-MBRL algorithm that reduces gradient error by adapting the model-based horizon to avoid stiff dynamics. Empirical findings reveal that AHAC outperforms MFRL baselines, attaining 40% more reward across a set of locomotion tasks and efficiently scaling to high-dimensional control environments with improved wall-clock-time efficiency. adaptive-horizon-actor-critic.github.io

Robots are often built from standardized assemblies, (e.g. arms, legs, or fingers), but each robot must be trained from scratch to control all the actuators of all the parts together. In this paper we demonstrate a new approach that takes a single robot and its controller as input and produces a set of modular controllers for each of these assemblies such that when a new robot is built from the same parts, its control can be quickly learned by reusing the modular controllers. We achieve this with a framework called MeMo which learns (Me)aningful, (Mo)dular controllers. Specifically, we propose a novel modularity objective to learn an appropriate division of labor among the modules. We demonstrate that this objective can be optimized simultaneously with standard behavior cloning loss via noise injection. We benchmark our framework in locomotion and grasping environments on simple to complex robot morphology transfer. We also show that the modules help in task transfer. On both structure and task transfer, MeMo achieves improved training efficiency to graph neural network and Transformer baselines.

Cloth simulation has wide applications in computer animation, garment design, and robot-assisted dressing. This work presents a differentiable cloth simulator whose additional gradient information facilitates cloth-related applications. Our differentiable simulator extends a state-of-the-art cloth simulator based on Projective Dynamics (PD) and with dry frictional contact [Ly et al. 2020 ]. We draw inspiration from previous work [Du et al. 2021 ] to propose a fast and novel method for deriving gradients in PD-based cloth simulation with dry frictional contact. Furthermore, we conduct a comprehensive analysis and evaluation of the usefulness of gradients in contact-rich cloth simulation. Finally, we demonstrate the efficacy of our simulator in a number of downstream applications, including system identification, trajectory optimization for assisted dressing, closed-loop control, inverse design, and real-to-sim transfer. We observe a substantial speedup obtained from using our gradient information in solving most of these applications.

Autonomous underwater vehicles (AUVs) are spe-cialized robots that are commonly used for seafloor surveying and ocean water sampling. Computational design approaches have emerged to reduce the effort required to design both individual AUVs as well as fleets. As the number and scale of underwater missions increases beyond the capabilities of a single vehicle, fleet level design will become more important. Depending on the mission, the optimal fleet may consist of multiple distinct types of AUVs designed to a variety of specifications. Moreover, the AUVs may differ in both continuous parameters (such as battery capacity) and discrete parameters (such as number and model of thrusters). In this work, we present a computational pipeline for designing these heterogeneous AUV fleets. Using a novel shape design space based on a graph grammar and deformation cages, we can express a variety of AUV architectures with different topologies, component selections, and dimensions. We search this space using a combination of discrete graph search and gradient-based continuous optimization, enabled by a differentiable AUV simulator. Finally, we formulate heterogeneous fleet design as a modified knapsack problem, and solve it using an efficient backtracking-based algorithm. We evaluate our pipeline on a simulated mission with nonuniform design requirements-surveying a section of seafloor with varying depth-and show that the best heterogeneous fleet outperforms the best fleet composed of a single vehicle type.

Traditional robotic manipulator design methods require extensive, time-consuming, and manual trial and error to produce a viable design. During this process, engineers often spend their time redesigning or reshaping components as they discover better topologies for the robotic manipula-tor. Tactile sensors, while useful, often complicate the design due to their bulky form factor. We propose an integrated design pipeline to streamline the design and manufacturing of robotic manipulators with knitted, glove-like tactile sensors. The proposed pipeline allows a designer to assemble a collection of modular, open-source components by applying predefined graph grammar rules. The end result is an intuitive design paradigm that allows the creation of new virtual designs of manipulators in a matter of minutes. Our framework allows the designer to fine-tune the manipulator's shape through cage-based geometry deformation. Finally, the designer can select surfaces for adding tactile sensing. Once the manipulator design is finished, the program will automatically generate 3D printing and knitting files for manufacturing. We demonstrate the utility of this pipeline by creating four custom manipulators tested on real-world tasks: screwing in a wing nut, pouring water from a bottle, picking up an egg, and cutting paper with scissors.

Unmanned aerial vehicles (UAVs) have broad applications including disaster response, transportation, photography, and mapping. A significant bottleneck in the development of UAVs is the limited availability of automatic tools for task-specific co-design of a UAV's shape and controller. The development of such tools is particularly challenging as UAVs can take many forms, including fixed-wing planes, radial copters, and hybrid topologies, with each class of topology showing different advantages. In this work, we present a computational design pipeline for UAVs based on a graph grammar that can search across a wide range of topologies. Graphs generated by the grammar encode different topologies and component selections, while continuous parameters encode the dimensions and properties of each component. We further augment the shape representation with deformation cages, which allow expressing a variety of wing shapes. Each UAV design is associated with an LQR controller with tunable continuous parameters. To search over this complex discrete and continuous design space, we develop a hybrid algorithm that combines discrete graph search strategies and gradient-based continuous optimization methods using a differentiable UAV simulator. We evaluate our pipeline on a set of simulated flight tasks requiring dynamic motions, showing that it discovers novel UAV designs that outperform canonical UAVs typically made by engineers.

Enabling additive manufacturing to employ a wide range of novel, functional materials can be a major boost to this technology. However, making such materials printable requires painstaking trial-and-error by an expert operator, as they typically tend to exhibit peculiar rheological or hysteresis properties. Even in the case of successfully finding the process parameters, there is no guarantee of print-to-print consistency due to material differences between batches. These challenges make closed-loop feedback an attractive option where the process parameters are adjusted on-the-fly. There are several challenges for designing an efficient controller: the deposition parameters are complex and highly coupled, artifacts occur after long time horizons, simulating the deposition is computationally costly, and learning on hardware is intractable. In this work, we demonstrate the feasibility of learning a closed-loop control policy for additive manufacturing using reinforcement learning. We show that approximate, but efficient, numerical simulation is sufficient as long as it allows learning the behavioral patterns of deposition that translate to real-world experiences. In combination with reinforcement learning, our model can be used to discover control policies that outperform baseline controllers. Furthermore, the recovered policies have a minimal sim-to-real gap. We showcase this by applying our control policy in-vivo on a single-layer printer using low and high viscosity materials.

Assembly planning is the core of automating product assembly, maintenance, and recycling for modern industrial manufacturing. Despite its importance and long history of research, planning for mechanical assemblies when given the final assembled state remains a challenging problem. This is due to the complexity of dealing with arbitrary 3D shapes and the highly constrained motion required for real-world assemblies. In this work, we propose a novel method to efficiently plan physically plausible assembly motion and sequences for real-world assemblies. Our method leverages the assembly-by-disassembly principle and physics-based simulation to efficiently explore a reduced search space. To evaluate the generality of our method, we define a large-scale dataset consisting of thousands of physically valid industrial assemblies with a variety of assembly motions required. Our experiments on this new benchmark demonstrate we achieve a state-of-the-art success rate and the highest computational efficiency compared to other baseline algorithms. Our method also generalizes to rotational assemblies (e.g., screws and puzzles) and solves 80-part assemblies within several minutes.

We present methods for co-designing rigid robots over control and morphology (including discrete topology) over multiple objectives. Previous work has addressed problems in single-objective robot co-design or multi-objective control. However, the joint multi-objective co-design problem is extremely important for generating capable, versatile, algorithmically designed robots. In this work, we present Multi-Objective Graph Heuristic Search, which extends a single-objective graph heuristic search from previous work to enable a highly efficient multi-objective search in a combinatorial design topology space. Core to this approach, we introduce a new universal, multiobjective heuristic function based on graph neural networks that is able to effectively leverage learned information between different task trade-offs. We demonstrate our approach on six combinations of seven terrestrial locomotion and design tasks, including one three-objective example. We compare the captured Pareto fronts across different methods and demonstrate that our multi-objective graph heuristic search quantitatively and qualitatively outperforms other techniques.

Both the design and control of a robot play equally important roles in its task performance. However, while optimal control is well studied in the machine learning and robotics community, less attention is placed on finding the optimal robot design. This is mainly because co-optimizing design and control in robotics is characterized as a challenging problem, and more importantly, a comprehensive evaluation benchmark for co-optimization does not exist. In this paper, we propose Evolution Gym, the first large-scale benchmark for co-optimizing the design and control of soft robots. In our benchmark, each robot is composed of different types of voxels (e.g., soft, rigid, actuators), resulting in a modular and expressive robot design space. Our benchmark environments span a wide range of tasks, including locomotion on various types of terrains and manipulation. Furthermore, we develop several robot co-evolution algorithms by combining state-of-the-art design optimization methods and deep reinforcement learning techniques. Evaluating the algorithms on our benchmark platform, we observe robots exhibiting increasingly complex behaviors as evolution progresses, with the best evolved designs solving many of our proposed tasks. Additionally, even though robot designs are evolved autonomously from scratch without prior knowledge, they often grow to resemble existing natural creatures while outperforming hand-designed robots. Nevertheless, all tested algorithms fail to find robots that succeed in our hardest environments. This suggests that more advanced algorithms are required to explore the high-dimensional design space and evolve increasingly intelligent robots -- an area of research in which we hope Evolution Gym will accelerate progress. Our website with code, environments, documentation, and tutorials is available at http://evogym.csail.mit.edu/.

Many real-world control problems involve conflicting objectives where we desire a dense and high-quality set of control policies that are optimal for different objective preferences (called Pareto-optimal). While extensive research in multi-objective reinforcement learning (MORL) has been conducted to tackle such problems, multi-objective optimization for complex continuous robot control is still under-explored. In this work, we propose an efficient evolutionary learning algorithm to find the Pareto set approximation for continuous robot control problems, by extending a state-of-the-art RL algorithm and presenting a novel prediction model to guide the learning process. In addition to efficiently discovering the individual policies on the Pareto front, we construct a continuous set of Pareto-optimal solutions by Pareto analysis and interpolation. Furthermore, we design seven multi-objective RL environments with continuous action space, which is the first benchmark platform to evaluate MORL algorithms on various robot control problems. We test the previous methods on the proposed benchmark problems, and the experiments show that our approach is able to find a much denser and higher-quality set of Pareto policies than the existing algorithms.

We present RoboGrammar , a fully automated approach for generating optimized robot structures to traverse given terrains. In this framework, we represent each robot design as a graph, and use a graph grammar to express possible arrangements of physical robot assemblies. Each robot design can then be expressed as a sequence of grammar rules. Using only a small set of rules our grammar can describe hundreds of thousands of possible robot designs. The construction of the grammar limits the design space to designs that can be fabricated. For a given input terrain, the design space is searched to find the top performing robots and their corresponding controllers. We introduce Graph Heuristic Search - a novel method for efficient search of combinatorial design spaces. In Graph Heuristic Search, we explore the design space while simultaneously learning a function that maps incomplete designs (e.g., nodes in the combinatorial search tree) to the best performance values that can be achieved by expanding these incomplete designs. Graph Heuristic Search prioritizes exploration of the most promising branches of the design space. To test our method we optimize robots for a number of challenging and varied terrains. We demonstrate that RoboGrammar can successfully generate nontrivial robots that are optimized for a single terrain or a combination of terrains.

Hybrid unmanned aerial vehicles (UAV) combine advantages of multicopters and fixed-wing planes: vertical take-off, landing, and low energy use. However, hybrid UAVs are rarely used because controller design is challenging due to its complex, mixed dynamics. In this paper, we propose a method to automate this design process by training a mode-free, model-agnostic neural network controller for hybrid UAVs. We present a neural network controller design with a novel error convolution input trained by reinforcement learning. Our controller exhibits two key features: First, it does not distinguish among flying modes, and the same controller structure can be used for copters with various dynamics. Second, our controller works for real models without any additional parameter tuning process, closing the gap between virtual simulation and real fabrication. We demonstrate the efficacy of the proposed controller both in simulation and in our custom-built hybrid UAVs (Figure 1, 8). The experiments show that the controller is robust to exploit the complex dynamics when both rotors and wings are active in flight tests.

Point cloud segmentation is a fundamental problem. Due to the complexity of real-world scenes and the limitations of 3D scanners, interactive segmentation is currently the only way to cope with all kinds of point clouds. However, interactively segmenting complex and large-scale scenes is very time-consuming. In this paper, we present a novel interactive system for segmenting point cloud scenes. Our system automatically suggests a series of camera views, in which users can conveniently specify segmentation guidance. In this way, users may focus on specifying segmentation hints instead of manually searching for desirable views of unsegmented objects, thus significantly reducing user effort. To achieve this, we introduce a novel view preference model, which is based on a set of dedicated view attributes, with weights learned from a user study. We also introduce support relations for both graph-cut-based segmentation and finding similar objects. Our experiments show that our segmentation technique helps users quickly segment various types of scenes, outperforming alternative methods.

Computer Aided Design (CAD) is a multi-billion dollar industry used by almost every mechanical engineer in the world to create practically every existing manufactured shape. CAD models are not only widely available but also extremely useful in the growing field of fabrication-oriented design because they are parametric by construction and capture the engineer's design intent, including manufacturability. Harnessing this data, however, is challenging, because generating the geometry for a given parameter value requires time-consuming computations. Furthermore, the resulting meshes have different combinatorics, making the mesh data inherently discontinuous with respect to parameter adjustments. In our work, we address these challenges and develop tools that allow interactive exploration and optimization of parametric CAD data. To achieve interactive rates, we use precomputation on an adaptively sampled grid and propose a novel scheme for interpolating in this domain where each sample is a mesh with different combinatorics. Specifically, we extract partial correspondences from CAD representations for local mesh morphing and propose a novel interpolation method for adaptive grids that is both continuous/smooth and local (i.e., the influence of each sample is constrained to the local regions where mesh morphing can be computed). We show examples of how our method can be used to interactively visualize and optimize objects with a variety of physical properties.