Large Language Models (LLMs) work surprisingly well for some complex reasoning problems via chain-of-thought (CoT) or tree-of-thought (ToT), but the underlying reasons remain unclear. We seek to understand the performance of these methods by conducting experimental case studies and linking the outcomes to sample and computational complexity in machine learning. We found that if problems can be decomposed into a sequence of reasoning steps and learning to predict the next step has a low sample and computational complexity, explicitly outlining the reasoning chain with all necessary information for predicting the next step may improve performance. Conversely, for problems where predicting the next step is computationally hard, adopting ToT may yield better reasoning outcomes than attempting to formulate a short reasoning chain.
The advantages of pre-trained large language models (LLMs) are apparent in a variety of language processing tasks. But can a language model's knowledge be further harnessed to effectively disambiguate objects and navigate decision-making challenges within the realm of robotics? Our study reveals the LLM's aptitude for solving complex decision making challenges that are often previously modeled by Partially Observable Markov Decision Processes (POMDPs). A pivotal focus of our research is the object disambiguation capability of LLMs. We detail the integration of an LLM into a tabletop environment disambiguation task, a decision making problem where the robot's task is to discern and retrieve a user's desired object from an arbitrarily large and complex cluster of objects. Despite multiple query attempts with zero-shot prompt engineering (details can be found in the Appendix), the LLM struggled to inquire about features not explicitly provided in the scene description. In response, we have developed a few-shot prompt engineering system to improve the LLM's ability to pose disambiguating queries. The result is a model capable of both using given features when they are available and inferring new relevant features when necessary, to successfully generate and navigate down a precise decision tree to the correct object--even when faced with identical options.
This work addresses the problem of long-horizon task planning with the Large Language Model (LLM) in an open-world household environment. Existing works fail to explicitly track key objects and attributes, leading to erroneous decisions in long-horizon tasks, or rely on highly engineered state features and feedback, which is not generalizable. We propose a novel, expandable state representation that provides continuous expansion and updating of object attributes from the LLM's inherent capabilities for context understanding and historical action reasoning. Our proposed representation maintains a comprehensive record of an object's attributes and changes, enabling robust retrospective summary of the sequence of actions leading to the current state. This allows enhanced context understanding for decision-making in task planning. We validate our model through experiments across simulated and real-world task planning scenarios, demonstrating significant improvements over baseline methods in a variety of tasks requiring long-horizon state tracking and reasoning.
The data-driven approach to robot control has been gathering pace rapidly, yet generalization to unseen task domains remains a critical challenge. We argue that the key to generalization is representations that are (i) rich enough to capture all task-relevant information and (ii) invariant to superfluous variability between the training and the test domains. We experimentally study such a representation -- containing both depth and semantic information -- for visual navigation and show that it enables a control policy trained entirely in simulated indoor scenes to generalize to diverse real-world environments, both indoors and outdoors. Further, we show that our representation reduces the A-distance between the training and test domains, improving the generalization error bound as a result. Our proposed approach is scalable: the learned policy improves continuously, as the foundation models that it exploits absorb more diverse data during pre-training.
Teaching physical skills to humans requires one-on-one interaction between the teacher and the learner. With a shortage of human teachers, such a teaching mode faces the challenge of scaling up. Robots, with their replicable nature and physical capabilities, offer a solution. In this work, we present TeachingBot, a robotic system designed for teaching handwriting to human learners. We tackle two primary challenges in this teaching task: the adaptation to each learner's unique style and the creation of an engaging learning experience. TeachingBot captures the learner's style using a probabilistic learning approach based on the learner's handwriting. Then, based on the learned style, it provides physical guidance to human learners with variable impedance to make the learning experience engaging. Results from human-subject experiments based on 15 human subjects support the effectiveness of TeachingBot, demonstrating improved human learning outcomes compared to baseline methods. Additionally, we illustrate how TeachingBot customizes its teaching approach for individual learners, leading to enhanced overall engagement and effectiveness.
Vague objectives in many real-life scenarios pose long-standing challenges for robotics, as defining rules, rewards, or constraints for optimization is difficult. Tasks like tidying a messy table may appear simple for humans, but articulating the criteria for tidiness is complex due to the ambiguity and flexibility in commonsense reasoning. Recent advancement in Large Language Models (LLMs) offers us an opportunity to reason over these vague objectives: learned from extensive human data, LLMs capture meaningful common sense about human behavior. However, as LLMs are trained solely on language input, they may struggle with robotic tasks due to their limited capacity to account for perception and low-level controls. In this work, we propose a simple approach to solve the task of table tidying, an example of robotic tasks with vague objectives. Specifically, the task of tidying a table involves not just clustering objects by type and functionality for semantic tidiness but also considering spatial-visual relations of objects for a visually pleasing arrangement, termed as visual tidiness. We propose to learn a lightweight, image-based tidiness score function to ground the semantically tidy policy of LLMs to achieve visual tidiness. We innovatively train the tidiness score using synthetic data gathered using random walks from a few tidy configurations. Such trajectories naturally encode the order of tidiness, thereby eliminating the need for laborious and expensive human demonstrations. Our empirical results show that our pipeline can be applied to unseen objects and complex 3D arrangements.
Inverse reinforcement learning (IRL) algorithms often rely on (forward) reinforcement learning or planning over a given time horizon to compute an approximately optimal policy for a hypothesized reward function and then match this policy with expert demonstrations. The time horizon plays a critical role in determining both the accuracy of reward estimate and the computational efficiency of IRL algorithms. Interestingly, an effective time horizon shorter than the ground-truth value often produces better results faster. This work formally analyzes this phenomenon and provides an explanation: the time horizon controls the complexity of an induced policy class and mitigates overfitting with limited data. This analysis leads to a principled choice of the effective horizon for IRL. It also prompts us to reexamine the classic IRL formulation: it is more natural to learn jointly the reward and the effective horizon together rather than the reward alone with a given horizon. Our experimental results confirm the theoretical analysis.
In the autonomous driving system, trajectory prediction plays a vital role in ensuring safety and facilitating smooth navigation. However, we observe a substantial discrepancy between the accuracy of predictors on fixed datasets and their driving performance when used in downstream tasks. This discrepancy arises from two overlooked factors in the current evaluation protocols of trajectory prediction: 1) the dynamics gap between the dataset and real driving scenario; and 2) the computational efficiency of predictors. In real-world scenarios, prediction algorithms influence the behavior of autonomous vehicles, which, in turn, alter the behaviors of other agents on the road. This interaction results in predictor-specific dynamics that directly impact prediction results. As other agents' responses are predetermined on datasets, a significant dynamics gap arises between evaluations conducted on fixed datasets and actual driving scenarios. Furthermore, focusing solely on accuracy fails to address the demand for computational efficiency, which is critical for the real-time response required by the autonomous driving system. Therefore, in this paper, we demonstrate that an interactive, task-driven evaluation approach for trajectory prediction is crucial to reflect its efficacy for autonomous driving.
Natural language provides a natural interface for human communication, yet it is challenging for robots to comprehend due to its abstract nature and inherent ambiguity. Large language models (LLMs) contain commonsense knowledge that can help resolve language ambiguity and generate possible solutions to abstract specifications. While LLMs have shown promise as few-shot planning policies, their potential for planning complex tasks is not fully tapped. This paper shows that LLMs can be used as both the commonsense model of the world and the heuristic policy in search algorithms such as Monte Carlo Tree Search (MCTS). MCTS explores likely world states sampled from LLMs to facilitate better-reasoned decision-making. The commonsense policy from LLMs guides the search to relevant parts of the tree, substantially reducing the search complexity. We demonstrate the effectiveness of our method in daily task-planning experiments and highlight its advantages over using LLMs solely as policies.
Identifying internal parameters for planning is crucial to maximizing the performance of a planner. However, automatically tuning internal parameters which are conditioned on the problem instance is especially challenging. A recent line of work focuses on learning planning parameter generators, but lack a consistent problem definition and software framework. This work proposes the unified planner optimization problem (POP) formulation, along with the Open Planner Optimization Framework (OPOF), a highly extensible software framework to specify and to solve these problems in a reusable manner.