Domain randomization in reinforcement learning is an established technique for increasing the robustness of control policies learned in simulation. By randomizing properties of the environment during training, the learned policy can be robust to uncertainty along the randomized dimensions. While the environment distribution is typically specified by hand, in this paper we investigate the problem of automatically discovering this sampling distribution via entropy-regularized reward maximization of a neural sampling distribution in the form of a normalizing flow. We show that this architecture is more flexible and results in better robustness than existing approaches to learning simple parameterized sampling distributions. We demonstrate that these policies can be used to learn robust policies for contact-rich assembly tasks. Additionally, we explore how these sampling distributions, in combination with a privileged value function, can be used for out-of-distribution detection in the context of an uncertainty-aware multi-step manipulation planner.

Despite significant technological advancements, the process of programming robots for adaptive assembly remains labor-intensive, demanding expertise in multiple domains and often resulting in task-specific, inflexible code. This work explores the potential of Large Language Models (LLMs), like ChatGPT, to automate this process, leveraging their ability to understand natural language instructions, generalize examples to new tasks, and write code. In this paper, we suggest how these abilities can be harnessed and applied to real-world challenges in the manufacturing industry. We present a novel system that uses ChatGPT to automate the process of programming robots for adaptive assembly by decomposing complex tasks into simpler subtasks, generating robot control code, executing the code in a simulated workcell, and debugging syntax and control errors, such as collisions. We outline the architecture of this system and strategies for task decomposition and code generation. Finally, we demonstrate how our system can autonomously program robots for various assembly tasks in a real-world project.

The automated assembly of complex products requires a system that can automatically plan a physically feasible sequence of actions for assembling many parts together. In this paper, we present ASAP, a physics-based planning approach for automatically generating such a sequence for general-shaped assemblies. ASAP accounts for gravity to design a sequence where each sub-assembly is physically stable with a limited number of parts being held and a support surface. We apply efficient tree search algorithms to reduce the combinatorial complexity of determining such an assembly sequence. The search can be guided by either geometric heuristics or graph neural networks trained on data with simulation labels. Finally, we show the superior performance of ASAP at generating physically realistic assembly sequence plans on a large dataset of hundreds of complex product assemblies. We further demonstrate the applicability of ASAP on both simulation and real-world robotic setups. Project website: asap.csail.mit.edu

Industrial insertion tasks are often performed repetitively with parts that are subject to tight tolerances and prone to breakage. Learning an industrial insertion policy in real is challenging as the collision between the parts and the environment can cause slippage or breakage of the part. In this paper, we present a safe self-supervised method to learn a visuo-tactile insertion policy that is robust to grasp pose variations. The method reduces human input and collisions between the part and the receptacle. The method divides the insertion task into two phases. In the first align phase, a tactile-based grasp pose estimation model is learned to align the insertion part with the receptacle. In the second insert phase, a vision-based policy is learned to guide the part into the receptacle. The robot uses force-torque sensing to achieve a safe self-supervised data collection pipeline. Physical experiments on the USB insertion task from the NIST Assembly Taskboard suggest that the resulting policies can achieve 45/45 insertion successes on 45 different initial grasp poses, improving on two baselines: (1) a behavior cloning agent trained on 50 human insertion demonstrations (1/45) and (2) an online RL policy (TD3) trained in real (0/45).

We introduce a robotic assembly system that streamlines the design-to-make workflow for going from a CAD model of a product assembly to a fully programmed and adaptive assembly process. Our system captures (in the CAD tool) the intent of the assembly process for a specific robotic workcell and generates a recipe of task-level instructions. By integrating visual sensing with deep-learned perception models, the robots infer the necessary actions to assemble the design from the generated recipe. The perception models are trained directly from simulation, allowing the system to identify various parts based on CAD information. We demonstrate the system with a workcell of two robots to assemble interlocking 3D part designs. We first build and tune the assembly process in simulation, verifying the generated recipe. Finally, the real robotic workcell assembles the design using the same behavior.

Co-robots, i.e. robots that work close to people, will need to account for the preferences and expectations of their human co-workers in executing trajectories or actions. Consistent, legible and predictable trajectories are a key factor in making humans comfortable around robots. In this work, we take a data-driven approach towards designing robot trajectories that are more acceptable to human co-workers and observers. We use an online survey to ask people to rate multiple robot trajectories generated in a variety of environments. We compute a large set of features for each trajectory, also taking into account environment information. We use a combination of the features and the survey ratings to learn a classifier that predicts the rating for a new trajectory based on the learned human-observer preferences. The classifier also helps identify and highlight the most important features that influence people's ratings of the trajectories. Finally, we discuss how a data-driven approach using the results of this analysis can be used to help design better trajectories that are more acceptable to people.

Opening and navigating through doors remains a challenging problem, particularly in cluttered environments and for spring-loaded doors. Passing through doors, especially spring-loaded doors, requires making and breaking contacts with the door and preventing the door from closing while passing through. In this work, we present a planning framework that handles non-spring and spring-loaded doors, in cluttered or confined workspaces, planning the approach to the door, pushing or pulling it open, and passing through. Because the problem is solved in a combined search space, the planner yields an overall least-cost path. The planner is able to insert a transition between robot-door contacts at any point along the plan. We utilize a compact graph-based representation of the problem to keep planning times low. We precompute the force workspace of the end-effectors to eliminate checks against joint torque limits at plan time. We have validated our solution in both simulation and real-world experiments on the PR2 mobile manipulation platform; the robot is able to successfully open a variety of spring-loaded and non-spring-loaded doors by pushing and pulling.

Most prior techniques for proximity computations are designed for synthetic models and assume exact geometric representations. However, real robots construct representations of the environment using their sensors, and the generated representations are more cluttered and less precise than synthetic models. Furthermore, this sensor data is updated at high frequency. In this paper, we present new collision- and distance-query algorithms, which can efficiently handle large amounts of point cloud sensor data received at real-time rates. We present two novel techniques to accelerate the computation of broad-phase data structures: 1) we present a progressive technique that incrementally computes a high-quality dynamic AABB tree for fast culling, and 2) we directly use an octree representation of the point cloud data as a proximity data structure. We assign a probability value to each leaf node of the tree, and the algorithm computes the nodes corresponding to high collision probability. In practice, our new approaches can be an order of magnitude faster than previous methods. We demonstrate the performance of the new methods on both synthetic data and on sensor data collected using a Kinect™ for motion planning for a mobile manipulator robot.

Robots operating in real world environments need to find motion plans quickly. Robot motion should also be efficient and, when operating among people, predictable. Minimizing a cost function, e.g. path length, can produce short, reasonable paths. Anytime planners are ideal for this since they find an initial solution quickly and then improve solution quality as time permits. In previous work, we introduced the concept of Experience Graphs, which allow search-based planners to find paths with bounded sub-optimality quickly by reusing parts of previous paths where relevant. Here we extend planning with Experience Graphs to work in an anytime fashion so a first solution is found quickly using prior experience. As time allows, the dependence on this experience is reduced in order to produce closer to optimal solutions. We also demonstrate how Experience Graphs provide a new way of approaching incremental planning as they naturally reuse information when the environment, the starting configuration of the robot or the goal configuration change. Experimentally, we demonstrate the anytime and incremental properties of our algorithm on mobile manipulation tasks in both simulation and on a real PR2 robot.

Dual-arm manipulation is an increasingly important skill for robots operating in home, retail and industrial environments. Dual-arm manipulation is especially essential for tasks involving large objects which are harder to grasp and manipulate using a single arm. In this work, we address dual-arm manipulation of objects in indoor environments. We are particularly focused on tasks that involve an upright orientation constraint on the grasped object. Such constraints are often present in human environments, e.g. when manipulating a tray of food or a container with fluids. In this paper, we present a search-based approach that is capable of planning dual-arm motions, often within one second, in cluttered environments while adhering to the orientation constraints. Our approach systematically constructs a graph in task space and generates motions that are consistent across runs with similar start/goal configurations and are low-cost. These motions come with guarantees on completeness and bounds on the suboptimality with respect to the graph that encodes the planning problem. For many problems, the consistency of the generated motions is important as it helps make the actions of the robot more predictable for a human interacting with the robot.

Collision-free navigation in cluttered environments is essential for any mobile manipulation system. Traditional navigation systems have relied on a 2D grid map projected from a 3D representation for efficiency. This approach, however, prevents navigation close to objects in situations where projected 3D configurations are in collision within the 2D grid map even if actually no collision occurs in the 3D environment. Accordingly, when using such a 2D representation for planning paths of a mobile manipulation robot, the number of planning problems which can be solved is limited and suboptimal robot paths may result. We present a fast, integrated approach to solve path planning in 3D using a combination of an efficient octree-based representation of the 3D world and an anytime search-based motion planner. Our approach utilizes a combination of multi-layered 2D and 3D representations to improve planning speed, allowing the generation of almost real-time plans with bounded sub-optimality. We present extensive experimental results with the two-armed mobile manipulation robot PR2 carrying large objects in a highly cluttered environment. Using our approach, the robot is able to efficiently plan and execute trajectories while transporting objects, thereby often moving through demanding, narrow passageways.

We present a new collision and proximity library that integrates several techniques for fast and accurate collision checking and proximity computation. Our library is based on hierarchical representations and designed to perform multiple proximity queries on different model representations. The set of queries includes discrete collision detection, continuous collision detection, separation distance computation and penetration depth estimation. The input models may correspond to triangulated rigid or deformable models and articulated models. Moreover, FCL can perform probabilistic collision checking between noisy point clouds that are captured using cameras or LIDAR sensors. The main benefit of FCL lies in the fact that it provides a unified interface that can be used by various applications. Furthermore, its flexible architecture makes it easier to implement new algorithms within this framework. The runtime performance of the library is comparable to state of the art collision and proximity algorithms. We demonstrate its performance on synthetic datasets as well as motion planning and grasping computations performed using a two-armed mobile manipulation robot.

Randomized planners, search-based planners, potential-field approaches and trajectory optimization based motion planners are just some of the types of approaches that have been developed for motion planning. Given a motion planning problem, choosing the appropriate algorithm to use is a daunting task even for experts since there has been relatively little effort in comparing the plans generated by the different approaches, for different problems. In this paper, we present a set of benchmarks and the associated infrastructure for comparing different types of motion planning approaches and algorithms. The benchmarks are specifically designed for robotics and include typical indoor human environments. We present example motion planning problems for single arm tasks. Our infrastructure is designed to be easily extensible to allow for the addition of new planning approaches, new robots, new environments and new metrics. We present results comparing the performance of several motion planning algorithms to validate the use of these benchmarks.

Robots executing practical tasks in real environments are often subject to multiple constraints. These constraints include orientation constraints: e.g., keeping a glass of water upright, torque constraints: e.g., not exceeding the torque limits for an arm lifting heavy objects, visibility constraints: e.g., keeping an object in view while moving a robot arm, etc. Rejection sampling, Jacobian projection techniques and optimization-based approaches are just some of the methods that have been used to address such constraints while computing motion plans for robots performing manipulation tasks. In this work, we present an approach to handling certain types of constraints in a manner that significantly increases the efficiency of existing methods. Our approach focuses on the sampling step of a motion planner. We implement this step as the drawing of samples from a set that has been computed in advance instead of the direct sampling of constraints. We show how our approach can be applied to different constraints: orientation constraints on the end-effector of an arm, visibility constraints and dual-arm constraints. We present simulated results to validate our method, comparing it to approaches that use direct sampling of constraints.

In this paper, we develop an online motion planning approach which learns from its planning episodes (experiences) a graph, an Experience Graph. On the theoretical side, we show that planning with Experience graphs is complete and provides bounds on suboptimality with respect to the graph that represents the original planning problem. Experimentally, we show in simulations and on a physical robot that our approach is particularly suitable for higher-dimensional motion planning tasks such as planning for two armed mobile manipulation.

In this paper, we present a search-based motion planning algorithm for manipulation that handles the high dimensionality of the problem and minimizes the limitations associated with employing a strict set of pre-defined actions. Our approach employs a set of adaptive motion primitives comprised of static motions with variable dimensionality and on-the-fly motions generated by two analytical solvers. This method results in a slimmer, multi-dimensional lattice and offers the ability to satisfy goal constraints with precision. To validate our approach, we used a 7DOF manipulator to perform experiments on a real mobile manipulation platform (Willow Garage's PR2). Our results demonstrate the effectiveness of the planner in efficiently navigating cluttered spaces; the method generates consistent, low-cost motion trajectories, and guarantees the search is complete with bounds on the suboptimality of the solution.

We present a new approach to motion planning using a stochastic trajectory optimization framework. The approach relies on generating noisy trajectories to explore the space around an initial (possibly infeasible) trajectory, which are then combined to produced an updated trajectory with lower cost. A cost function based on a combination of obstacle and smoothness cost is optimized in each iteration. No gradient information is required for the particular optimization algorithm that we use and so general costs for which derivatives may not be available (e.g. costs corresponding to constraints and motor torques) can be included in the cost function. We demonstrate the approach both in simulation and on a mobile manipulation system for unconstrained and constrained tasks. We experimentally show that the stochastic nature of STOMP allows it to overcome local minima that gradient-based methods like CHOMP can get stuck in.

Learning complex motor skills for real world tasks is a hard problem in robotic manipulation that often requires painstaking manual tuning and design by a human expert. In this work, we present a Reinforcement Learning based approach to acquiring new motor skills from demonstration. Our approach allows the robot to learn fine manipulation skills and significantly improve its success rate and skill level starting from a possibly coarse demonstration. Our approach aims to incorporate task domain knowledge, where appropriate, by working in a space consistent with the constraints of a specific task. In addition, we also present an approach to using sensor feedback to learn a predictive model of the task outcome. This allows our system to learn the proprioceptive sensor feedback needed to monitor subsequent executions of the task online and abort execution in the event of predicted failure. We illustrate our approach using two example tasks executed with the PR2 dual-arm robot: a straight and accurate pool stroke and a box flipping task using two chopsticks as tools.

Robust navigation in cluttered environments has been well addressed for mobile robotic platforms, but the problem of navigating with a moveable object like a cart has not been widely examined. In this work, we present a planning and control approach to navigation of a humanoid robot while pushing a cart. We show how immediate information about the environment can be integrated into this approach to achieve safer navigation in the presence of dynamic obstacles. We demonstrate the robustness of our approach through long-running experiments with the PR2 mobile manipulation robot in a typical indoor office environment, where the robot faced narrow and high-traffic passageways with very limited clearance.

Tactile information is valuable in determining properties of objects that are inaccessible from visual perception. In this paper, we present a tactile perception strategy that allows a mobile robot with tactile sensors in its gripper to measure a generic set of tactile features while manipulating an object. We propose a switching velocity-force controller that grasps an object safely and reveals, at the same time, its deformation properties. By gently rolling the object, the robot can extract additional information about the contents of the object. As an application, we show that a robot can use these features to distinguish the internal state of bottles and cans-purely from tactile sensing-from a small training set. The robot can distinguish open from closed bottles and cans and full ones from empty ones. We also show how the high-frequency component in tactile information can be used to detect movement inside a container, e.g., in order to detect the presence of liquid. To prove that this is a hard recognition problem, we also conducted a comparative study with 17 human test subjects. The recognition rates of the human subjects were comparable with that of the robot.

We present a novel robotic grasp controller that allows a sensorized parallel jaw gripper to gently pick up and set down unknown objects once a grasp location has been selected. Our approach is inspired by the control scheme that humans employ for such actions, which is known to centrally depend on tactile sensation rather than vision or proprioception. Our controller processes measurements from the gripper's fingertip pressure arrays and hand-mounted accelerometer in real time to generate robotic tactile signals that are designed to mimic human SA-I, FA-I, and FA-II channels. These signals are combined into tactile event cues that drive the transitions between six discrete states in the grasp controller: Close, Load, Lift and Hold, Replace, Unload, and Open. The controller selects an appropriate initial grasping force, detects when an object is slipping from the grasp, increases the grasp force as needed, and judges when to release an object to set it down. We demonstrate the promise of our approach through implementation on the PR2 robotic platform, including grasp testing on a large number of real-world objects.

Computing a motion that enables a mobile manipulator to open a door is challenging because it requires tight coordination between the motions of the arm and the base. Hard-coding the motion, on the other hand, is infeasible since doors vary widely in their sizes and types, some doors are opened by pulling and others by pushing, and indoor spaces often contain obstacles that limit the freedom of the mobile manipulator and the degree to which the doors open up. In this paper, we show how to overcome the high-dimensionality of the planning problem by identifying a graph-based representation that is small enough for efficient planning yet rich enough to contain feasible motions that open doors. The use of graph search-based motion planning enables us to handle consistently the wide variance of conditions under which doors need to be open. We demonstrate our approach on the PR2 robot - a mobile manipulator with an omnidirectional base and a 7 degree of freedom arm. The robot was successful in opening a variety of doors both by pulling and pushing.

Tactile information is valuable in determining properties of objects that are inaccessible from visual perception. In this work, we present a tactile perception strategy that allows any mobile robot with tactile sensors in its gripper to measure a set of generic tactile features while grasping an object. We propose a hybrid velocity-force controller, that grasps an object safely and reveals at the same time its deformation properties. As an application, we show that a robot can use these features to distinguish the open/closed and fill state of bottles and cans - purely from tactile sensing - from a small training set. To prove that this is a hard recognition problem, we also conducted a comperative study with 17 human test subjects. We found that the recognition rate of the human subjects were comparable to our robotic gripper.

Heuristic searches such as A* search are highly popular means of finding least-cost plans due to their generality, strong theoretical guarantees on completeness and optimality and simplicity in the implementation. In planning for robotic manipulation however, these techniques are commonly thought of as impractical due to the high-dimensionality of the planning problem. In this paper, we present a heuristic search-based manipulation planner that does deal effectively with the high-dimensionality of the problem. The planner achieves the required efficiency due to the following three factors: (a) its use of informative yet fast-to-compute heuristics; (b) its use of basic (small) motion primitives as atomic actions; and (c) its use of ARA* search which is an anytime heuristic search with provable bounds on solution suboptimality. Our experimental analysis on a real mobile manipulation platform with a 7-DOF robotic manipulator shows the ability of the planner to solve manipulation in cluttered spaces by generating consistent, low-cost motion trajectories while providing guarantees on completeness and bounds on suboptimality.

We describe an autonomous robotic system capable of navigating through an office environment, opening doors along the way, and plugging itself into electrical outlets to recharge as needed. We demonstrate through extensive experimentation that our robot executes these tasks reliably, without requiring any modification to the environment. We present robust detection algorithms for doors, door handles, and electrical plugs and sockets, combining vision and laser sensors. We show how to overcome the unavoidable shortcoming of perception by integrating compliant control into manipulation motions. We present a visual-differencing approach to high-precision plug-insertion that avoids the need for high-precision hand-eye calibration.

We present a novel combination of motion planning techniques to compute motion plans for robotic arms. We compute plans that move the arm as close as possible to the goal region using sampling-based planning and then switch to a trajectory optimization technique for the last few centimeters necessary to reach the goal region. This combination allows fast computation and safe execution of motion plans even when the goals are very close to objects in the environment. The system incorporates realtime sensory inputs and correctly deals with occlusions that can occur when robot body parts block the sensor view of the environment. The system is tested on a 7 degree-of-freedom robot arm with sensory input from a tilting laser scanner that provides 3D information about the environment.

Robotic grasping in unstructured environments requires the ability to select grasps for unknown objects and execute them while dealing with uncertainty due to sensor noise or calibration errors. In this work, we propose a simple but robust approach to grasp selection for unknown objects, and a reactive adjustment approach to deal with uncertainty in object location and shape. The grasp selection method uses 3D sensor data directly to determine a ranked set of grasps for objects in a scene, using heuristics based on both the overall shape of the object and its local features. The reactive grasping approach uses tactile feedback from fingertip sensors to execute a compliant robust grasp. We present experimental results to validate our approach by grasping a wide range of unknown objects. Our results show that reactive grasping can correct for a fair amount of uncertainty in the measured position or shape of the objects, and that our grasp selection approach is successful in grasping objects with a variety of shapes.

We present a search-based planning approach for controlling a quadrupedal robot over rough terrain. Given a start and goal position, we consider the problem of generating a complete joint trajectory that will result in the legged robot successfully moving from the start to the goal. We decompose the problem into two main phases: an initial global planning phase, which results in a footstep trajectory; and an execution phase, which dynamically generates a joint trajectory to best execute the footstep trajectory. We show how R* search can be employed to generate high-quality global plans in the high-dimensional space of footstep trajectories. Results show that the global plans coupled with the joint controller result in a system robust enough to deal with a variety of terrains.

This paper presents significant steps towards the online integration of 3D perception and manipulation for personal robotics applications. We propose a modular and distributed architecture, which seamlessly integrates the creation of 3D maps for collision detection and semantic annotations, with a real-time motion replanning framework. To validate our system, we present results obtained during a comprehensive mobile manipulation scenario, which includes the fusion of the above components with a higher level executive.

We present a novel method for the localization of a legged robot on known terrain using only proprioceptive sensors such as joint encoders and an inertial measurement unit. In contrast to other proprioceptive pose estimation techniques, this method allows for global localization (i.e., localization with large initial uncertainty) without the use of exteroceptive sensors. This is made possible by establishing a measurement model based on the feasibility of putative poses on known terrain given observed joint angles and attitude measurements. Results are shown that demonstrate that the method performs better than dead-reckoning, and is also able to perform global localization from large initial uncertainty

The TRIKKE is a three-wheeled, human-powered scooter that can be propelled by a combination of cyclic motion of its handlebar and swaying motion of the rider. This paper addresses the modeling, dynamics and control of the TRIKKE and the development of a robotic platform called the ROBOTRIKKE that is derived from similar principles. The TRIKKE can be modeled as a modified roller-racer with an unstable steering arrangement. The model of the TRIKKE reduces to the roller-racer [8] in the absence of this steering arrangement. We prove that under certain conditions on the geometric parameters of the system, the TRIKKE and roller-racer systems cannot be stopped after motion starting from rest using the steering control as the sole input. As a consequence, the ideal model is severely limited from the point of view of controllability. We demonstrate the validity of our model through comparison with experimental measurements on a small-scale robotic prototype of the TRIKKE. We present closed-loop control results for tracking (on average) a straight line trajectory using visual feedback from an overhead camera.

We present the design and gait generation for an experimental ROLLERBLADER. The ROLLERBLADER is a robot with a central platform mounted on omnidirectional casters and two 3 degree-of-freedom legs. A passive rollerblading wheel is attached to the end of each leg. The wheels give rise to nonholonomic constraints acting on the robot. The legs can be picked up and placed back on the ground allowing a combination of skating and walking gaits. We present two types of gaits for the robot. In the first gait, we allow the legs to be picked up and placed back on the ground while in the second, the wheels are constrained to stay on the ground at all tunes. Experimental gait results for a prototype robot are also presented.

We develop the dynamic model for a planar rollerblader. The robot consists of a rigid platform and two planar, two degree-of-freedom legs with in-line skates at the foot. The dynamic model consists of two unicycles coupled through the rigid body dynamics of the planar platform. We derive the Lagrangian reduction for the rollerblading robot. We show the generation of some simple gaits that allow the platform to move forward and rotate by using cyclic motions of the two legs.

Addresses the issue of developing a motion planning algorithm for a general class of modular mobile robots. A modular mobile robot is essentially a reconfigurable robot system in which locomotive modules such as legs, wheels, propellers, etc. can be attached at various locations on the body. The equations of motion for the robot are developed in terms of a set of generalized inputs related to the constraints and configurations of each individual module. The motion planning method developed in Lafferriere and Sussmann (1991) and extended in Goodwine and Burdick (1997) can then be modified to generate required trajectories for the modular robot. These methods generally lead to sequential motion plans-we also develop conditions under which two consecutive plans can be executed simultaneously.

This paper presents several new insights and experimental results on the problem of generating omnidirectional walking gaits for quadrupedal robots. For statically stable gaits, by placing some minor restrictions on how the leg motions are generated, we develop an easily computable classification of the best gait patterns (leg phasings) to be used for any given motion. This simple geometric construction allows easy generation of stable, omnidirectional walking gaits for any quadruped satisfying only some very weak assumptions about the location of the center of mass relative to the workspace of the legs. It is also well-suited to a modular approach in which the control of individual legs is decentralized. We also implement an omnidirectional dynamic trot gait using a similar approach. Transitions involving a change in the direction of motion of the robot are initiated at one of two fixed points of the foot placement curves for each leg in all gaits. Experimental data is presented showing the effectiveness of these techniques.