IKEA Furniture Assembly Environment for Long-Horizon Complex Manipulation Tasks
The IKEA Furniture Assembly Environment is one of the first benchmarks for testing and accelerating the automation of complex manipulation tasks. The environment is designed to advance reinforcement learning from simple toy tasks to complex tasks requiring both long-term planning and sophisticated low-level control. Our environment supports over 80 different furniture models, Sawyer and Baxter robot simulation, and domain randomization. The IKEA Furniture Assembly Environment is a testbed for methods aiming to solve complex manipulation tasks. The environment is publicly available at https://clvrai.com/furnitureREAD FULL TEXT VIEW PDF
Robotic assembly planning has the potential to profoundly change how
This paper explores the idea that skillful assembly is best represented ...
The FetchIt! Mobile Manipulation Challenge, held at the IEEE Internation...
The dm_control software package is a collection of Python libraries and ...
Complex object manipulation tasks often span over long sequences of
Despite recent successes of reinforcement learning (RL), it remains a
Low-latency telerobotics can enable more intricate surface tasks on
IKEA Furniture Assembly Environment for Long-Horizon Complex Manipulation Tasks
The ability to plan and manipulate physical objects is a necessity to use tools, build structures, and ultimately interact with the world in a meaningful way. Thus, solving long-horizon manipulation tasks has been an active challenge in robot learning. One main bottleneck is the lack of a "standardized" simulation environment for long-term physical tasks. Such a simulator for robot learning needs to have the following properties: long-horizon hierarchical tasks for planning, high-quality rendering for sim-to-real transfer, variable agents and dynamic domains.
Imagine that you are building IKEA furniture. It is not trivial to figure out how to assemble pieces into the final configuration given only a glimpse of fully constructed furniture. Specifically, it is not apparent from pieces on the floor which parts to choose for attachment and in what order. Hence, we need to dissect the final configuration and deduce the sequence of tasks necessary to build the furniture by comparing the current and final configuration. Moreover, connecting two parts requires complicated manipulation skills, such as accurate alignment of two attaching points and sophisticated force control to firmly attach them. Therefore, furniture assembly is a comprehensive task requiring reliable perception, high-level planning, and sophisticated control, making it a suitable benchmark for robot learning algorithms.
To this end, we introduce the IKEA Furniture Assembly Environment
as a new benchmark for complex autonomous manipulation skills. The IKEA Furniture Assembly Environment is a visually realistic environment that simulates the task of furniture assembly as a step toward autonomous manipulation. The furniture assembly task involves not only high-level 3D scene understanding and step-by-step planning but also sophisticated low-level control. Figure1 shows examples of our rendered environments.
A variety of research problems could be investigated with this new environment, namely perception, planning, and control. For perception, the environment could be used to solve 3D object detection, pose estimation, instance segmentation, scene graph generation, and shape estimation problems. For robotic control, the environment is suitable for testing multi-agent reinforcement learning, hierarchical reinforcement learning, model-based reinforcement learning, imitation learning, and sim-to-real algorithms for long-term complex manipulation tasks.
The IKEA Furniture Assembly Environment simulates over 80 furniture models. The environment supports multiple agents such as Cursor, Baxter and Sawyer robots. To test generalization, the environment supports domain randomization in the furniture, physics, lighting, textures, and more factors of variation found in the real world. We also outline in the conclusion promising future directions for complex manipulation tasks.
To advance reinforcement learning from simple, videogame-esque tasks to complex and realistic tasks, our environment features long-horizon, hierarchical tasks, realistic rendering, domain randomization, and accurate robotic simulation. The furniture assembly task is composed of (1) selecting two compatible parts, (2) grasping these two parts, (3) aligning the attachable spots (namely connectors) of the parts, and (4) connecting the parts as seen in Figure 2. The furniture assembly task can be accomplished by repeating this process until all parts are assembled. We first discuss various research topics that can utilize our environment, and then present the design and implementation of the environment.
The IKEA Furniture Assembly environment simulates complex long-term robotic manipulation tasks with 80+ furniture models and different robots. Moreover, the environment features realistic rendering, configurable scene (e.g., lighting, texture, color), and diverse ground truth labels, such as object pose, shape, instance segmentation mask, depth map, and scene graph. Hence, the environment can be used not only for learning planning and control but also for learning perception as following:
The environment can generate an abundant amount of data with diverse labels, such as object pose, shape, instance segmentation mask, depth map, part configurations, and scene graph. This synthetic data can be used to tackle many computer vision problems, such as object pose estimation, semantic segmentation, scene graph understanding, shape estimation.
Control: The furniture assembly task requires sophisticated manipulation skills. We support impedance control, inverse kinematics for task space control, as well as keyboard control for humans. Thus, the proposed environment can be used as a challenging benchmark for reinforcement learning and imitation learning methods. Bi-manual or multi-agent manipulation of furniture is another interesting direction.
Planning: The furniture assembly consists of multiple steps of assembling two furniture parts and thus has a long horizon. To tackle this long-horizon task, an agent should learn to plan using approaches from model-based planning and hierarchical reinforcement learning.
To cover several challenges including 3D alignment of various shapes of objects and long-horizon robotic manipulation, we developed a novel 3D environment that supports assembling IKEA furniture models using MuJoCo (Todorov et al., 2012) as a physics simulator and Unity3D game engine as a renderer. MuJoCo provides fast and accurate physics simulation, while Unity3d has superior texture and lighting configuration. To exploit the strengths of both frameworks, we use MuJoCo as the underlying physics engine and Unity3d as the renderer. To enable the robotic simulation of Sawyer and Baxter arms in MuJoCo, we use the Robosuite framework (Fan et al., 2018). We also use DoorGym (Urakami et al., 2019) to integrate Unity rendering engine with MuJoCo.
The simulation environment follows the OpenAI Gym (Brockman et al., 2016) protocol where an environment takes an action as input and takes a step. Robotic arms can move around the environment and interact with furniture parts. In addition to actions for robotic arm moves, our environment has a connect action. During each step, the environment checks all pairs of connectors. Currently, we only support one-to-one mapping between connectors and plan to implement many-to-many mappings for identical parts. If a pair of connectors have matching IDs and the connectors are within a positional and angular distance threshold, the pair of connectors are flagged as attachable. Specifically, we check the euclidean distance between the connector coordinates (Equation (1
)), the cosine distance between the connector up vectors (Equation (2)), and the cosine distance between the connector forward vectors (Equation (3)). If the connect action is activated, then the attachable parts are connected using the MuJoCo weld mechanism. The thresholds for distance and angle are configurable. For full details, refer to the _is_aligned function in the furniture.py file.
Currently, the environment provides a sparse reward for successfully connecting two furniture parts. Users can easily implement their own reward functions and use the rewards for training an agent. For example, the Baxter-Block-Pick task implemented in the furniture_baxter_block.py file computes the reward function composed of 10 different components including the distance between the gripper and the target object, angle of the gripper, height of the target object, and gripper’s state.
Our environment provides a simulation of the furniture assembly of over 80 different furniture models as shown in Figure 4. We define the furniture as a combination of parts and connectors.
The furniture parts are modeled by following the IKEA’s official user’s manuals with a minor simplification in small details such as carving and screws. The furniture models are created using 3D modeling tool Rhino111https://www.rhino3d.com/ and each furniture part is converted to a separate 3D mesh file in a format of STL. Given an STL 3D mesh file, MuJoCo can represent the mesh as a physical object. Each furniture part is represented as one or more meshes. Concave furniture parts need to be represented with multiple meshes, as MuJoCo only supports convex meshes for collision detection. To enable collision of a concave mesh, we split a concave mesh into multiple convex meshes with STL editing software or use MuJoCo’s primitive meshes (e.g., box, cylinder, sphere) only for the collision detection.
A pair of connectors define the connection information between two furniture parts. The connectors are located on the parts and serve as areas of attachment. On a given part, we parameterize connectors with their ID, size, position, and orientation to the part. All furniture parts and connectors are defined through the MuJoCo XML file. Refer to the XML documentation in the codebase for information and examples on the schema.
To handle part connection, we define in the XML the connector constraints (e.g. Part A can connect to Part B in a certain pose) with the MuJoCo weld equality constraint, which allows for welding together two given parts. Once all parts and their connector constraints are specified, the furniture model is complete and can be loaded into the MuJoCo simulator for simulation.
The environment supports a variety of agents for interacting with furniture. Currently, three agents are available: Cursor, Sawyer, and Baxter, as illustrated in Figure 5. Their action spaces and observation spaces are fully configurable to fit a variety of problem settings. The observation space can consist of agent state (e.g., joint positions and velocities), environmental state (e.g., coordinates and rotation of furniture parts), and a camera observation. Aside from RGB images, the environment also supports object segmentation masks and depth camera observations as seen in Figure 3. In general, the action space consists of movement, selection, and attachment primitives.
The Cursor agent abstracts away the low-level object grasping problem. Composed of two floating and collision-free cursors in the environment, the agent can move the cursors in the (x,y,z)-axes, and hold parts that are encapsulated by the cubes. Parts that are held by the cursor can be rotated on the (x,y,z)-axes. Movement and rotation can either be applied continuously or in discrete steps depending on the configuration. This agent is suitable for methods that focus on the planning and reasoning portion of the furniture assembly problem by abstracting away the low-level object manipulation problem. The internal state contains the (x,y,z)-coordinates of the cursor.
The Sawyer and Baxter robots are available for simulation in the environment. Sawyer uses 7 DoF robotic arm, while the Baxter robot has two 7 DoF arms. They can be controlled either through impedance control and task space control (end-effector pose). For task space control, we use the PyBullet library (Coumans, 2015) to calculate inverse kinematics for the end-effector control. Unlike the cursor agent, object selection is not abstracted, and objects need to be grasped realistically by the grippers using contact forces. The internal state contains the joint angles and gripping status of the robot. An egocentric or third-person viewpoint can be selected for the camera. RGB, segmentation, and depth modes are available for the camera.
To reduce the reality gap (Jakobi et al., 1995), all the furniture models are created following the IKEA’s official user’s manuals with a minor simplification in small details such as carving and screws. For generalization of the learned skills, the environment should provide enough variability in furniture compositions, visual appearances, object shapes, and physical properties. For example, the environment will contain a diverse set of furniture including chair, table, cabinet, bookcase, desk, shelf, and tv unit (see Figure 4). For given furniture, the environment can randomly spawn subsets of the furniture parts, and randomly initialize their positions and orientation them in the scene to increase generalization. In addition to random part selection and placement, the environment can also randomize physics such as gravity and friction to add more variation in the task. The environment will also support the randomization of visual properties like lighting, background colors, texture, and more. Figure 6 illustrates examples of domain randomization.
Our environment contains several limitations. First, connection process of two furniture parts are not physically simulated due to the difficulty of physics simulation of screwing. Instead of simulating screws, the environment allows two parts to be connected if two connectors of parts are close and well-aligned as described in Equation (4). Next, identical furniture pieces are considered differently and one connector can only be connected to the predefined furniture part, unlike the real world where the identical pieces can be used interchangeably. In the next release, we plan to allow a furniture part replacing another identical part.
The followings are possible future extensions:
3D motion devices: VR controllers and 3D mouses can be used to teleoperate robots to collect human demonstration. The data can be used for training robot agents.
Additional robot support: more diverse robots, such as Fetch, UR, and higher DoF grippers, can be trained using the proposed environment.
Realistic part attachment: the connection mechanism can be more realistic by implementing peg insertion, screwing, nailing mechanisms.
Tool use: an agent can learn to use a screwdriver or a hammer for assembling furniture.
Multi-agent assembly: multiple robots collaborate to assemble complex furniture.
Instructions and demonstrations: language and visual instructions or demonstrations can be used to guide an agent to solve a complex manipulation task.
Reinforcement learning has made rapid progress with the advent of standardized and simulated environments. Most progress has been made in game environments, such as Atari (Bellemare et al., 2013), OpenAI gym (Brockman et al., 2016), VizDoom (Kempka et al., 2016), and StarCraft2 (Vinyals et al., 2017). Recently, many simulated environments have been introduced in diverse applications, such as autonomous driving (Shah et al., 2017; Dosovitskiy et al., 2017), indoor navigation (Kolve et al., 2017; Puig et al., 2018), continuous control (Tassa et al., 2018; Ghosh et al., 2018; Lee et al., 2019a), and recommendation systems (Rohde et al., 2018).
In robotics, most existing environments focus on short-term object manipulation tasks, such as picking and placing (Zhu et al., 2018; Lee et al., 2019b), in-hand dexterous manipulation (Andrychowicz et al., 2018; Rajeswaran et al., 2018), door opening (Urakami et al., 2019), and peg inserting (Chebotar et al., 2018). Recent advancements in simulators have made a push towards more complex and realistic tasks. Fan et al. (2018) takes a step towards a comprehensive manipulation simulator by offering a variety of manipulation tasks. However, the tasks, consisting of lifting, stacking, pick and place, are still limited to primitive skills.
Composite manipulation tasks such as block stacking (Duan et al., 2017; Xu et al., 2018) and ball serving (Lee et al., 2019a) have also been proposed. However, these tasks have small variations in shapes and physical properties of objects. Metaworld and RLBench (Yu et al., 2019; James et al., 2019) offer environments with diverse but simple manipulation tasks. In contrast, we propose a complex manipulation task with a hierarchical task structure and long horizon, furniture assembly, which requires long-term planning and generalizable skills for various shapes, textures, and materials of objects.
The bottleneck of transferring an agent trained in simulation to real-world robots is the reality gap between simulation and the real world (Jakobi et al., 1995). Recently, domain randomization technique (Sadeghi and Levine, 2017) has been proposed to reduce the reality gap by training the policy on multiple instances of data differing in textures, backgrounds, lighting and more. Once trained with enough variation, the policy should be able to generalize to the real world (Rajeswaran et al., 2016; Tobin et al., 2017; Sadeghi et al., 2017; Andrychowicz et al., 2018; Urakami et al., 2019)
. To enable transfer learning of the challenging furniture assembly task, the IKEA Furniture Assembly Environment supports diverse and configurable textures, backgrounds, and lighting. Moreover, we provide over 80 furniture models with 3 different robots for further domain randomization.
Solving the furniture assembly task requires the solving of multiple subproblems, such as perception, planning, and control. Each of these sub-problems are nontrivial and requires domain knowledge and supervision. For recognizing furniture models and their structures, IKEA 3D model dataset (Lim et al., 2013) provide labeled data about 3D structures of furniture models and part information. The IKEA furniture assembly video dataset (Han et al., 2017), which consists of 480,000 video frames of humans assembling a table from IKEA, can be used as demonstration data for imitation learning. Our proposed environment can be also used to generate synthetic training data for a wide range of 3D perception models with instance segmentation masks, depth image, and furniture part information under diverse furniture models and backgrounds. The expert demonstration videos and trajectories can be collected and used for imitation learning, similar to Roboturk (Mandlekar et al., 2018).
Mujoco (Todorov et al., 2012), Unity ML-Agents (Juliani et al., 2018), PyBullet (Coumans, 2015), and Dart (Lee et al., 2018) are some popular frameworks for physics simulation for reinforcement learning. We chose to use MuJoCo due to fast simulation and its widespread usage in the reinforcement learning community. MuJoCo also supports a Unity plugin for rendering MuJoCo scenes, enabling realistic lighting, scene, and background variations for domain randomization (Sadeghi and Levine, 2017).
In this paper, we propose the IKEA Furniture Assembly Environment as a novel benchmark for testing complex manipulation tasks. Furniture assembly is a complex manipulation task even for humans, requiring perception capabilities, high-level planning, and sophisticated low-level control. Therefore, it is well suited as a benchmark for control algorithms aiming to solve complex tasks.
We present several directions for future research that are necessary for completely automating furniture assembly.
Representation: Currently, many manipulation methods still require the pose of the object. This is difficult to obtain in the real world. On the other hand, images of the scene are easy to get but obtaining relevant features is difficult due to its high dimensionality. To solve furniture assembly in the real world, extracting object-centric representations from images is a promising direction. These representations may have geometric, relational, and semantic information (Hu et al., 2018; Zambaldi et al., 2019; Greff et al., 2017)
. Unsupervised object discovery with deep neural networks(Greff et al., 2017; Burgess et al., 2019) can be also helpful.
Planning: While reinforcement learning has seen success in manipulation tasks, the manipulation tasks are short in horizon and do not require complex planning. Furniture assembly on the other hand, requires long-term planning such as deciding the ordering of parts to assemble. Hierarchical reinforcement learning can segment complex tasks into simpler subtasks, which may be conducive towards furniture assembly (Sutton et al., 1999; Lee et al., 2019a). Model-based reinforcement learning could be another direction for tackling the long-horizon problem by giving the model a mechanism to predict the future before taking an action (Sutton, 1991).
: Dexterous manipulation is challenging due to agent and object complexity. Real world objects have irregular shapes, textures, friction, and other physical characteristics. Some dexterous robotic hands have between 24-30 degrees of freedom. Model free methods have seen success in controlling such dexterous hands(Andrychowicz et al., 2017; Rajeswaran et al., 2018) but gaining full control of robotic hands and grasping arbitrary objects is an open challenge. Multi-agent manipulation of furniture is also another promising direction.
Domain Knowledge: Instruction manuals (Andreas et al., 2017; Oh et al., 2017), programs (Xu et al., 2018; Sun et al., 2018), and video demonstrations (Lee et al., 2019b) can be used as domain knowledge to help solving complex tasks. The question of how to integrate additional supervision into an automated furniture assembly process is an interesting future direction.
The authors would like to thank Taehoon Kim and many members of the USC CLVR lab for helpful feedback and testing our environment. This project was partially funded by SKT.
Surreal: Open-source reinforcement learning framework and robot manipulation benchmark.In Conference on Robot Learning, pages 767–782, 2018.
Journal of Artificial Intelligence Research, 47:253–279, jun 2013.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8494–8502, 2018.
Neural expectation maximization.In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, pages 6691–6701, 2017.
International Conference on Machine Learning, pages 166–175, 2017.