Reinforcement learning (RL) agents often face challenges in balancing exploration and exploitation, particularly in environments where sparse or dense rewards bias learning. Biological systems, such as human toddlers, naturally navigate this balance by transitioning from free exploration with sparse rewards to goal-directed behavior guided by increasingly dense rewards. Inspired by this natural progression, we investigate the Toddler-Inspired Reward Transition in goal-oriented RL tasks. Our study focuses on transitioning from sparse to potential-based dense (S2D) rewards while preserving optimal strategies. Through experiments on dynamic robotic arm manipulation and egocentric 3D navigation tasks, we demonstrate that effective S2D reward transitions significantly enhance learning performance and sample efficiency. Additionally, using a Cross-Density Visualizer, we show that S2D transitions smooth the policy loss landscape, resulting in wider minima that improve generalization in RL models. In addition, we reinterpret Tolman's maze experiments, underscoring the critical role of early free exploratory learning in the context of S2D rewards.

Applying multi-agent reinforcement learning methods to realistic settings is challenging as it may require the agents to quickly adapt to unexpected situations that are rarely or never encountered in training. Recent methods for generalization to such out-of-distribution settings are limited to more specific, restricted instances of distribution shifts. To tackle adaptation to distribution shifts, we propose Unexpected Encoding Scheme, a novel decentralized multi-agent reinforcement learning algorithm where agents communicate "unexpectedness," the aspects of the environment that are surprising. In addition to a message yielded by the original reward-driven communication, each agent predicts the next observation based on previous experience, measures the discrepancy between the prediction and the actually encountered observation, and encodes this discrepancy as a message. Experiments on multi-robot warehouse environment support that our proposed method adapts robustly to dynamically changing training environments as well as out-of-distribution environment.

This paper explores how non-experts can teach robots desired skills in their environments. We argue that natural language is an intuitive and accessible interface for robot learning. To this end, we investigate two key aspects: (1) how non-experts collect robotic data using natural language supervision and (2) how pre-trained vision-language models learn end-to-end policies directly from this supervision. We propose a data collection framework that collects robot demonstrations based on natural language supervision (e.g., "move forward") and further augments these demonstrations. Next, we introduce a model that learns language-conditioned policies from natural language supervision called CLIP-RT. Our model employs pre-trained CLIP models and learns to predict actions represented in language via contrastive imitation learning. We first train CLIP-RT on large-scale robotic data and then enable it to learn desired skills using data collected from our framework. CLIP-RT shows strong capabilities in acquiring novel manipulation skills, outperforming the state-of-the-art model, OpenVLA (7B parameters), by 17% in average success rates, while using 7x fewer parameters (1B).

We propose the zero-shot Vision-and-Language Navigation with Collision Mitigation (VLN-CM), which takes these considerations. VLN-CM is composed of four modules and predicts the direction and distance of the next movement at each step. We utilize large foundation models for each modules. To select the direction, we use the Attention Spot Predictor (ASP), View Selector (VS), and Progress Monitor (PM). The ASP employs a Large Language Model (e.g. ChatGPT) to split navigation instructions into attention spots, which are objects or scenes at the location to move to (e.g. a yellow door). The VS selects from panorama images provided at 30-degree intervals the one that includes the attention spot, using CLIP similarity. We then choose the angle of the selected image as the direction to move in. The PM uses a rule-based approach to decide which attention spot to focus on next, among multiple spots derived from the instructions. If the similarity between the current attention spot and the visual observations decreases consecutively at each step, the PM determines that the agent has passed the current spot and moves on to the next one. For selecting the distance to move, we employed the Open Map Predictor (OMP). The OMP uses panorama depth information to predict an occupancy mask. We then selected a collision-free distance in the predicted direction based on the occupancy mask. We evaluated our method using the validation data of VLN-CE. Our approach showed better performance than several baseline methods, and the OPM was effective in mitigating collisions for the agent.

Causal dynamics learning has recently emerged as a promising approach to enhancing robustness in reinforcement learning (RL). Typically, the goal is to build a dynamics model that makes predictions based on the causal relationships among the entities. Despite the fact that causal connections often manifest only under certain contexts, existing approaches overlook such fine-grained relationships and lack a detailed understanding of the dynamics. In this work, we propose a novel dynamics model that infers fine-grained causal structures and employs them for prediction, leading to improved robustness in RL. The key idea is to jointly learn the dynamics model with a discrete latent variable that quantizes the state-action space into subgroups. This leads to recognizing meaningful context that displays sparse dependencies, where causal structures are learned for each subgroup throughout the training. Experimental results demonstrate the robustness of our method to unseen states and locally spurious correlations in downstream tasks where fine-grained causal reasoning is crucial. We further illustrate the effectiveness of our subgroup-based approach with quantization in discovering fine-grained causal relationships compared to prior methods.

Monte Carlo Tree Search (MCTS) has showcased its efficacy across a broad spectrum of decision-making problems. However, its performance often degrades under vast combinatorial action space, especially where an action is composed of multiple sub-actions. In this work, we propose an action abstraction based on the compositional structure between a state and sub-actions for improving the efficiency of MCTS under a factored action space. Our method learns a latent dynamics model with an auxiliary network that captures sub-actions relevant to the transition on the current state, which we call state-conditioned action abstraction. Notably, it infers such compositional relationships from high-dimensional observations without the known environment model. During the tree traversal, our method constructs the state-conditioned action abstraction for each node on-the-fly, reducing the search space by discarding the exploration of redundant sub-actions. Experimental results demonstrate the superior sample efficiency of our method compared to vanilla MuZero, which suffers from expansive action space.

Conditional independence provides a way to understand causal relationships among the variables of interest. An underlying system may exhibit more fine-grained causal relationships especially between a variable and its parents, which will be called the local independence relationships. One of the most widely studied local relationships is Context-Specific Independence (CSI), which holds in a specific assignment of conditioned variables. However, its applicability is often limited since it does not allow continuous variables: data conditioned to the specific value of a continuous variable contains few instances, if not none, making it infeasible to test independence. In this work, we define and characterize the local independence relationship that holds in a specific set of joint assignments of parental variables, which we call context-set specific independence (CSSI). We then provide a canonical representation of CSSI and prove its fundamental properties. Based on our theoretical findings, we cast the problem of discovering multiple CSSI relationships in a system as finding a partition of the joint outcome space. Finally, we propose a novel method, coined neural contextual decomposition (NCD), which learns such partition by imposing each set to induce CSSI via modeling a conditional distribution. We empirically demonstrate that the proposed method successfully discovers the ground truth local independence relationships in both synthetic dataset and complex system reflecting the real-world physical dynamics.

Human perception involves discerning complex multi-object scenes into time-static object appearance (i.e., size, shape, color) and time-varying object motion (i.e., location, velocity, acceleration). This innate ability to unconsciously understand the environment is the motivation behind the success of dynamics modeling. Object-centric representations have emerged as a promising tool for dynamics prediction, yet they primarily focus on the objects' appearance, often overlooking other crucial attributes. In this paper, we propose Object-Centric Kinematics (OCK), a framework for dynamics prediction leveraging objectcentric representations. Our model utilizes a novel component named object kinematics, which comprises low-level structured states of objects' position, velocity, and acceleration. The object kinematics are obtained via either implicit or explicit approaches, enabling comprehensive spatiotemporal object reasoning, and integrated through various transformer mechanisms, facilitating effective object-centric dynamics modeling. Our model demonstrates superior performance when handling objects and backgrounds in complex scenes characterized by a wide range of object attributes and dynamic movements. Moreover, our model demonstrates generalization capabilities across diverse synthetic environments, highlighting its potential for broad applicability in vision-related tasks.

Embodied Instruction Following (EIF) is the task of executing natural language instructions by navigating and interacting with objects in 3D environments. One of the primary challenges in EIF is compositional task planning, which is often addressed with supervised or in-context learning with labeled data. To this end, we introduce the Socratic Planner, the first zero-shot planning method that infers without the need for any training data. Socratic Planner first decomposes the instructions into substructural information of the task through self-questioning and answering, translating it into a high-level plan, i.e., a sequence of subgoals. Subgoals are executed sequentially, with our visually grounded re-planning mechanism adjusting plans dynamically through a dense visual feedback. We also introduce an evaluation metric of high-level plans, RelaxedHLP, for a more comprehensive evaluation. Experiments demonstrate the effectiveness of the Socratic Planner, achieving competitive performance on both zero-shot and few-shot task planning in the ALFRED benchmark, particularly excelling in tasks requiring higher-dimensional inference. Additionally, a precise adjustments in the plan were achieved by incorporating environmental visual information.

Vision-and-Language Navigation (VLN) agents navigate to a destination using natural language instructions and the visual information they observe. Existing methods for training VLN agents presuppose fixed datasets, leading to a significant limitation: the introduction of new environments necessitates retraining with previously encountered environments to preserve their knowledge. This makes it difficult to train VLN agents that operate in the ever-changing real world. To address this limitation, we present the Continual Vision-and-Language Navigation (CVLN) paradigm, designed to evaluate agents trained through a continual learning process. For the training and evaluation of CVLN agents, we re-arrange existing VLN datasets to propose two datasets: CVLN-I, focused on navigation via initial-instruction interpretation, and CVLN-D, aimed at navigation through dialogue with other agents. Furthermore, we propose two novel rehearsal-based methods for CVLN, Perplexity Replay (PerpR) and Episodic Self-Replay (ESR). PerpR prioritizes replaying challenging episodes based on action perplexity, while ESR replays previously predicted action logits to preserve learned behaviors. We demonstrate the effectiveness of the proposed methods on CVLN through extensive experiments.