Practice Makes Perfect: Planning to Learn Skill Parameter Policies

One promising approach towards effective robot decision making in complex, long-horizon tasks is to sequence together parameterized skills. We consider a setting where a robot is initially equipped with (1) a library of parameterized skills, (2) an AI planner for sequencing together the skills given a goal, and (3) a very general prior distribution for selecting skill parameters. Once deployed, the robot should rapidly and autonomously learn to improve its performance by specializing its skill parameter selection policy to the particular objects, goals, and constraints in its environment. In this work, we focus on the active learning problem of choosing which skills to practice to maximize expected future task success. We propose that the robot should estimate the competence of each skill, extrapolate the competence (asking: "how much would the competence improve through practice?"), and situate the skill in the task distribution through competence-aware planning. This approach is implemented within a fully autonomous system where the robot repeatedly plans, practices, and learns without any environment resets. Through experiments in simulation, we find that our approach learns effective parameter policies more sample-efficiently than several baselines. Experiments in the real-world demonstrate our approach's ability to handle noise from perception and control and improve the robot's ability to solve two long-horizon mobile-manipulation tasks after a few hours of autonomous practice.

E(2)-Equivariant Graph Planning for Navigation

Learning for robot navigation presents a critical and challenging task. The scarcity and costliness of real-world datasets necessitate efficient learning approaches. In this letter, we exploit Euclidean symmetry in planning for 2D navigation, which originates from Euclidean transformations between reference frames and enables parameter sharing. To address the challenges of unstructured environments, we formulate the navigation problem as planning on a geometric graph and develop an equivariant message passing network to perform value iteration. Furthermore, to handle multi-camera input, we propose a learnable equivariant layer to lift features to a desired space. We conduct comprehensive evaluations across five diverse tasks encompassing structured and unstructured environments, along with maps of known and unknown, given point goals or semantic goals. Our experiments confirm the substantial benefits on training efficiency, stability, and generalization.

Can Euclidean Symmetry be Leveraged in Reinforcement Learning and Planning?

In robotic tasks, changes in reference frames typically do not influence the underlying physical properties of the system, which has been known as invariance of physical laws.These changes, which preserve distance, encompass isometric transformations such as translations, rotations, and reflections, collectively known as the Euclidean group. In this work, we delve into the design of improved learning algorithms for reinforcement learning and planning tasks that possess Euclidean group symmetry. We put forth a theory on that unify prior work on discrete and continuous symmetry in reinforcement learning, planning, and optimal control. Algorithm side, we further extend the 2D path planning with value-based planning to continuous MDPs and propose a pipeline for constructing equivariant sampling-based planning algorithms. Our work is substantiated with empirical evidence and illustrated through examples that explain the benefits of equivariance to Euclidean symmetry in tackling natural control problems.

Equivariant Single View Pose Prediction Via Induced and Restricted Representations

Learning about the three-dimensional world from two-dimensional images is a fundamental problem in computer vision. An ideal neural network architecture for such tasks would leverage the fact that objects can be rotated and translated in three dimensions to make predictions about novel images. However, imposing SO(3)-equivariance on two-dimensional inputs is difficult because the group of three-dimensional rotations does not have a natural action on the two-dimensional plane. Specifically, it is possible that an element of SO(3) will rotate an image out of plane. We show that an algorithm that learns a three-dimensional representation of the world from two dimensional images must satisfy certain geometric consistency properties which we formulate as SO(2)-equivariance constraints. We use the induced and restricted representations of SO(2) on SO(3) to construct and classify architectures which satisfy these geometric consistency constraints. We prove that any architecture which respects said consistency constraints can be realized as an instance of our construction. We show that three previously proposed neural architectures for 3D pose prediction are special cases of our construction. We propose a new algorithm that is a learnable generalization of previously considered methods. We test our architecture on three pose predictions task and achieve SOTA results on both the PASCAL3D+ and SYMSOL pose estimation tasks.

Scaling up and Stabilizing Differentiable Planning with Implicit Differentiation

Differentiable planning promises end-to-end differentiability and adaptivity. However, an issue prevents it from scaling up to larger-scale problems: they need to differentiate through forward iteration layers to compute gradients, which couples forward computation and backpropagation, and needs to balance forward planner performance and computational cost of the backward pass. To alleviate this issue, we propose to differentiate through the Bellman fixed-point equation to decouple forward and backward passes for Value Iteration Network and its variants, which enables constant backward cost (in planning horizon) and flexible forward budget and helps scale up to large tasks. We study the convergence stability, scalability, and efficiency of the proposed implicit version of VIN and its variants and demonstrate their superiorities on a range of planning tasks: 2D navigation, visual navigation, and 2-DOF manipulation in configuration space and workspace.

Integrating Symmetry into Differentiable Planning

We study how group symmetry helps improve data efficiency and generalization for end-to-end differentiable planning algorithms, specifically on 2D robotic path planning problems: navigation and manipulation. We first formalize the idea from Value Iteration Networks (VINs) on using convolutional networks for path planning, because it avoids explicitly constructing equivalence classes and enable end-to-end planning. We then show that value iteration can always be represented as some convolutional form for (2D) path planning, and name the resulting paradigm Symmetric Planner (SymPlan). In implementation, we use steerable convolution networks to incorporate symmetry. Our algorithms on navigation and manipulation, with given or learned maps, improve training efficiency and generalization performance by large margins over non-equivariant counterparts, VIN and GPPN.

Toward Compositional Generalization in Object-Oriented World Modeling

Compositional generalization is a critical ability in learning and decision-making. We focus on the setting of reinforcement learning in object-oriented environments to study compositional generalization in world modeling. We (1) formalize the compositional generalization problem with an algebraic approach and (2) study how a world model can achieve that. We introduce a conceptual environment, Object Library, and two instances, and deploy a principled pipeline to measure the generalization ability. Motivated by the formulation, we analyze several methods with exact} or no compositional generalization ability using our framework, and design a differentiable approach, Homomorphic Object-oriented World Model (HOWM), that achieves approximate but more efficient compositional generalization.

Learning Symmetric Embeddings for Equivariant World Models

Incorporating symmetries can lead to highly data-efficient and generalizable models by defining equivalence classes of data samples related by transformations. However, characterizing how transformations act on input data is often difficult, limiting the applicability of equivariant models. We propose learning symmetric embedding networks (SENs) that encode an input space (e.g. images), where we do not know the effect of transformations (e.g. rotations), to a feature space that transforms in a known manner under these operations. This network can be trained end-to-end with an equivariant task network to learn an explicitly symmetric representation. We validate this approach in the context of equivariant transition models with 3 distinct forms of symmetry. Our experiments demonstrate that SENs facilitate the application of equivariant networks to data with complex symmetry representations. Moreover, doing so can yield improvements in accuracy and generalization relative to both fully-equivariant and non-equivariant baselines.

Deep Imitation Learning for Bimanual Robotic Manipulation

We present a deep imitation learning framework for robotic bimanual manipulation in a continuous state-action space. Imitation learning has been effectively utilized in mimicking bimanual manipulation movements, but generalizing the movement to objects in different locations has not been explored. We hypothesize that to precisely generalize the learned behavior relative to an object's location requires modeling relational information in the environment. To achieve this, we designed a method that (i) uses a multi-model framework to decomposes complex dynamics into elemental movement primitives, and (ii) parameterizes each primitive using a recurrent graph neural network to capture interactions. Our model is a deep, hierarchical, modular architecture with a high-level planner that learns to compose primitives sequentially and a low-level controller which integrates primitive dynamics modules and inverse kinematics control. We demonstrate the effectiveness using several simulated bimanual robotic manipulation tasks. Compared to models based on previous imitation learning studies, our model generalizes better and achieves higher success rates in the simulated tasks.