We introduce a new method that extracts knowledge from a large language model (LLM) to produce object-level plans, which describe high-level changes to object state, and uses them to bootstrap task and motion planning (TAMP). Existing work uses LLMs to directly output task plans or generate goals in representations like PDDL. However, these methods fall short because they rely on the LLM to do the actual planning or output a hard-to-satisfy goal. Our approach instead extracts knowledge from an LLM in the form of plan schemas as an object-level representation called functional object-oriented networks (FOON), from which we automatically generate PDDL subgoals. Our method markedly outperforms alternative planning strategies in completing several pick-and-place tasks in simulation.††Project Website: https://davidpaulius.github.io/olpllm/

The distributional hypothesis states that the meaning of a concept is defined through the contexts it occurs in. In practice, often word co-occurrence and proximity are analyzed in text corpora for a given word to obtain a real-valued semantic word vector, which is taken to (at least partially) encode the meaning of this word. Here we transfer this idea from text to images, where pre-assigned labels of other objects or activations of convolutional neural networks serve as context. We propose a simple algorithm that extracts and processes object contexts from an image database and yields semantic vectors for objects. We show empirically that these representations exhibit on par performance with state-of-the-art distributional models over a set of conventional objects. For this we employ well-known word benchmarks in addition to a newly proposed object-centric benchmark.

Function approximation in online, incremental, reinforcement learning needs to deal with two fundamental problems: biased sampling and nonstationarity. In this kind of task, biased sampling occurs because samples are obtained from specific trajectories dictated by the dynamics of the environment and are usually concentrated in particular convergence regions, which in the long term tend to dominate the approximation in the less sampled regions. The nonstationarity comes from the recursive nature of the estimations typical of temporal difference methods. This nonstationarity has a local profile, varying not only along the learning process but also along different regions of the state space. We propose to deal with these problems using an estimation of the probability density of samples represented with a gaussian mixture model. To deal with the nonstationarity problem, we use the common approach of introducing a forgetting factor in the updating formula. However, instead of using the same forgetting factor for the whole domain, we make it dependent on the local density of samples, which we use to estimate the nonstationarity of the function at any given input point. To address the biased sampling problem, the forgetting factor applied to each mixture component is modulated according to the new information provided in the updating, rather than forgetting depending only on time, thus avoiding undesired distortions of the approximation in less sampled regions.

In this work we address the problem of finding replacements of missing objects that are needed for the execution of human-like manipulation tasks. This is a usual problem that is easily solved by humans provided their natural knowledge to find object substitutions: using a knife as a screwdriver or a book as a cutting board. On the other hand, in robotic applications, objects required in the task should be included in advance in the problem definition. If any of these objects is missing from the scenario, the conventional approach is to manually redefine the problem according to the available objects in the scene. In this work we propose an automatic way of finding object substitutions for the execution of manipulation tasks. The approach uses a logic-based planner to generate a plan from a prototypical problem definition and searches for replacements in the scene when some of the objects involved in the plan are missing. This is done by means of a repository of objects and attributes with roles, which is used to identify the affordances of the unknown objects in the scene. Planning actions are grounded using a novel approach that encodes the semantic structure of manipulation actions. The system was evaluated in a KUKA arm platform for the task of preparing a salad with successful results.

Robotic applications in human environments are usually implemented using a cognitive architecture that integrates techniques of different levels of abstraction, ranging from artificial intelligence techniques for making decisions at a symbolic level to robotic techniques for grounding symbolic actions. In this work we address the problem of simultaneous learning at different levels of abstractions in such an architecture. This problem is important since human environments are highly variable, and many unexpected situations may arise during the execution of a task. The usual approach under this circumstance is to train each level individually to learn how to deal with the new situations. However, this approach is limited since it implies long task interruptions every time a new situation needs to be learned. We propose an architecture where learning takes place simultaneously at all the levels of abstraction. To achieve this, we devise a method that permits higher levels to guide the learning at the levels below for the correct execution of the task. The architecture is instantiated with a logic-based planner and an online planning operator learner, at the highest level, and with online reinforcement learning units that learn action policies for the grounding of the symbolic actions, at the lowest one. A human teacher is involved in the decision-making loop to facilitate learning. The framework is tested in a physically realistic simulation of the Sokoban game.

In the online version of the EM algorithm introduced by Sato and Ishii ( 2000 ), a time-dependent discount factor is introduced for forgetting the effect of the old estimated values obtained with an earlier, inaccurate estimator. In their approach, forgetting is uniformly applied to the estimators of each mixture component depending exclusively on time, irrespective of the weight attributed to each unit for the observed sample. This causes an excessive forgetting in the less frequently sampled regions. To address this problem, we propose a modification of the algorithm that involves a weight-dependent forgetting, different for each mixture component, in which old observations are forgotten according to the actual weight of the new samples used to replace older values. A comparison of the time-dependent versus the weight-dependent approach shows that the latter improves the accuracy of the approximation and exhibits much greater stability.

Recent approaches to Reinforcement Learning (RL) with function approximation include Neural Fitted Q Iteration and the use of Gaussian Processes. They belong to the class of fitted value iteration algorithms, which use a set of support points to fit the value-function in a batch iterative process. These techniques make efficient use of a reduced number of samples by reusing them as needed, and are appropriate for applications where the cost of experiencing a new sample is higher than storing and reusing it, but this is at the expense of increasing the computational effort, since these algorithms are not incremental. On the other hand, non-parametric models for function approximation, like Gaussian Processes, are preferred against parametric ones, due to their greater flexibility. A further advantage of using Gaussian Processes for function approximation is that they allow to quantify the uncertainty of the estimation at each point. In this paper, we propose a new approach for RL in continuous domains based on Probability Density Estimations. Our method combines the best features of the previous methods: it is non-parametric and provides an estimation of the variance of the approximated function at any point of the domain. In addition, our method is simple, incremental, and computationally efficient. All these features make this approach more appealing than Gaussian Processes and fitted value iteration algorithms in general.

A central problem in reinforcement learning is how to deal with large state and action spaces. When the problem domain presents intrinsic symmetries, exploiting them can be key to achieve good performance. We analyze the gains that can be effectively achieved by exploiting different kinds of symmetries, and the effect of combining them, in a test case: the stand-up and stabilization of an inverted pendulum.

Real world robot applications have to cope with large variations in the operating conditions due to the variability and unpredictability of the environment and its interaction with the robot. Performing an adequate control using conventional control techniques, that require the model of the plant and some knowledge about the influence of the environment, could be almost impossible. An alternative to traditional control techniques is to use an automatic learning system that uses previous experience to learn an adequate control policy. Learning by experience has been formalized in the field of reinforcement learning. But the application of reinforcement learning techniques in complex environments is only feasible when some generalization can be made in order to reduce the required amount of experience. This work presents an algorithm that performs a kind of generalization called categorization. This algorithm is able to perform efficient generalization of the observed situations, and learn accurate control policies in a short time without any previous knowledge of the plant and without the need of any kind of traditional control technique. Its performance is evaluated on the trajectory tracking control with simulated DC motors and compared with PID systems specifically tuned for the same problem.