In planning, stitching is an ability of algorithms to piece together sub-trajectories of data they are trained on to generate new and diverse behaviours. While stitching is historically a strength of offline reinforcement learning, recent generative behavioural cloning (BC) methods have also shown proficiency at stitching. However, the main factors behind this are poorly understood, hindering the development of new algorithms that can reliably stitch. Focusing on diffusion planners trained via BC, we find two properties are needed to compose: \emph{positional equivariance} and \emph{local receptiveness}. We use these two properties to explain architecture, data, and inference choices in existing generative BC methods based on diffusion planning, including replanning frequency, data augmentation, and data scaling. Experimental comparisions show that (1) while locality is more important than positional equivariance in creating a diffusion planner capable of composition, both are crucial (2) enabling these properties through relatively simple architecture choices can be competitive with more computationally expensive methods such as replanning or scaling data, and (3) simple inpainting-based guidance can guide architecturally compositional models to enable generalization in goal-conditioned settings.
Offline reinforcement learning (RL) looks at learning how to optimally solve tasks using a fixed dataset of interactions from the environment. Many off-policy algorithms developed for online learning struggle in the offline setting as they tend to over-estimate the behaviour of out of distributions actions. Existing offline RL algorithms adapt off-policy algorithms, employing techniques such as constraining the policy or modifying the value function to achieve good performance on individual datasets but struggle to adapt to different tasks or datasets of different qualities without tuning hyper-parameters. We introduce a policy switching technique that dynamically combines the behaviour of a pure off-policy RL agent, for improving behaviour, and a behavioural cloning (BC) agent, for staying close to the data. We achieve this by using a combination of epistemic uncertainty, quantified by our RL model, and a metric for aleatoric uncertainty extracted from the dataset. We show empirically that our policy switching technique can outperform not only the individual algorithms used in the switching process but also compete with state-of-the-art methods on numerous benchmarks. Our use of epistemic uncertainty for policy switching also allows us to naturally extend our method to the domain of offline to online fine-tuning allowing our model to adapt quickly and safely from online data, either matching or exceeding the performance of current methods that typically require additional modification or hyper-parameter fine-tuning.
The combination of behavioural cloning and neural networks has driven significant progress in robotic manipulation. As these algorithms may require a large number of demonstrations for each task of interest, they remain fundamentally inefficient in complex scenarios, in which finite datasets can hardly cover the state space. One of the remaining challenges is thus out-of-distribution (OOD) generalisation, i.e. the ability to predict correct actions for states with a low likelihood with respect to the state occupancy induced by the dataset. This issue is aggravated when the system to control is treated as a black-box, ignoring its physical properties. This work characterises widespread properties of robotic manipulation, specifically pose equivariance and locality. We investigate the effect of the choice of problem space on OOD performance of BC policies and how transformations arising from characteristic properties of manipulation could be employed for its improvement. We empirically demonstrate that these transformations allow behaviour cloning policies, using either standard MLP-based one-step action prediction or diffusion-based action-sequence prediction, to generalise better to OOD problem instances.
When interacting with each other, humans adjust their behavior based on perceived trust. However, to achieve similar adaptability, robots must accurately estimate human trust at sufficiently granular timescales during the human-robot collaboration task. A beta reputation is a popular way to formalize a mathematical estimation of human trust. However, it relies on binary performance, which updates trust estimations only after each task concludes. Additionally, manually crafting a reward function is the usual method of building a performance indicator, which is labor-intensive and time-consuming. These limitations prevent efficiently capturing continuous changes in trust at more granular timescales throughout the collaboration task. Therefore, this paper presents a new framework for the estimation of human trust using a beta reputation at fine-grained timescales. To achieve granularity in beta reputation, we utilize continuous reward values to update trust estimations at each timestep of a task. We construct a continuous reward function using maximum entropy optimization to eliminate the need for the laborious specification of a performance indicator. The proposed framework improves trust estimations by increasing accuracy, eliminating the need for manually crafting a reward function, and advancing toward developing more intelligent robots. The source code is publicly available. https://github.com/resuldagdanov/robot-learning-human-trust
Offline-to-online reinforcement learning (O2O RL) aims to obtain a continually improving policy as it interacts with the environment, while ensuring the initial policy behaviour is satisficing. This satisficing behaviour is necessary for robotic manipulation where random exploration can be costly due to catastrophic failures and time. O2O RL is especially compelling when we can only obtain a scarce amount of (potentially suboptimal) demonstrations$\unicode{x2014}$a scenario where behavioural cloning (BC) is known to suffer from distribution shift. Previous works have outlined the challenges in applying O2O RL algorithms under the image-based environments. In this work, we propose a novel O2O RL algorithm that can learn in a real-life image-based robotic vacuum grasping task with a small number of demonstrations where BC fails majority of the time. The proposed algorithm replaces the target network in off-policy actor-critic algorithms with a regularization technique inspired by neural tangent kernel. We demonstrate that the proposed algorithm can reach above 90\% success rate in under two hours of interaction time, with only 50 human demonstrations, while BC and existing commonly-used RL algorithms fail to achieve similar performance.
Vision-language navigation (VLN) requires an agent to execute actions following human instructions. Existing VLN models are optimized through expert demonstrations by supervised behavioural cloning or incorporating manual reward engineering. While straightforward, these efforts overlook the accumulation of errors in the Markov decision process, and struggle to match the distribution of the expert policy. Going beyond this, we propose an Energy-based Navigation Policy (ENP) to model the joint state-action distribution using an energy-based model. At each step, low energy values correspond to the state-action pairs that the expert is most likely to perform, and vice versa. Theoretically, the optimization objective is equivalent to minimizing the forward divergence between the occupancy measure of the expert and ours. Consequently, ENP learns to globally align with the expert policy by maximizing the likelihood of the actions and modeling the dynamics of the navigation states in a collaborative manner. With a variety of VLN architectures, ENP achieves promising performances on R2R, REVERIE, RxR, and R2R-CE, unleashing the power of existing VLN models.
Current methods to learn controllers for autonomous vehicles (AVs) focus on behavioural cloning. Being trained only on exact historic data, the resulting agents often generalize poorly to novel scenarios. Simulators provide the opportunity to go beyond offline datasets, but they are still treated as complicated black boxes, only used to update the global simulation state. As a result, these RL algorithms are slow, sample-inefficient, and prior-agnostic. In this work, we leverage a differentiable simulator and design an analytic policy gradients (APG) approach to training AV controllers on the large-scale Waymo Open Motion Dataset. Our proposed framework brings the differentiable simulator into an end-to-end training loop, where gradients of the environment dynamics serve as a useful prior to help the agent learn a more grounded policy. We combine this setup with a recurrent architecture that can efficiently propagate temporal information across long simulated trajectories. This APG method allows us to learn robust, accurate, and fast policies, while only requiring widely-available expert trajectories, instead of scarce expert actions. We compare to behavioural cloning and find significant improvements in performance and robustness to noise in the dynamics, as well as overall more intuitive human-like handling.
Social learning plays an important role in the development of human intelligence. As children, we imitate our parents' speech patterns until we are able to produce sounds; we learn from them praising us and scolding us; and as adults, we learn by working with others. In this work, we survey the degree to which this paradigm -- social learning -- has been mirrored in machine learning. In particular, since learning socially requires interacting with others, we are interested in how embodied agents can and have utilised these techniques. This is especially in light of the degree to which recent advances in natural language processing (NLP) enable us to perform new forms of social learning. We look at how behavioural cloning and next-token prediction mirror human imitation, how learning from human feedback mirrors human education, and how we can go further to enable fully communicative agents that learn from each other. We find that while individual social learning techniques have been used successfully, there has been little unifying work showing how to bring them together into socially embodied agents.
Some imitation learning methods combine behavioural cloning with self-supervision to infer actions from state pairs. However, most rely on a large number of expert trajectories to increase generalisation and human intervention to capture key aspects of the problem, such as domain constraints. In this paper, we propose Continuous Imitation Learning from Observation (CILO), a new method augmenting imitation learning with two important features: (i) exploration, allowing for more diverse state transitions, requiring less expert trajectories and resulting in fewer training iterations; and (ii) path signatures, allowing for automatic encoding of constraints, through the creation of non-parametric representations of agents and expert trajectories. We compared CILO with a baseline and two leading imitation learning methods in five environments. It had the best overall performance of all methods in all environments, outperforming the expert in two of them.
Often times in imitation learning (IL), the environment we collect expert demonstrations in and the environment we want to deploy our learned policy in aren't exactly the same (e.g. demonstrations collected in simulation but deployment in the real world). Compared to policy-centric approaches to IL like behavioural cloning, reward-centric approaches like inverse reinforcement learning (IRL) often better replicate expert behaviour in new environments. This transfer is usually performed by optimising the recovered reward under the dynamics of the target environment. However, (a) we find that modern deep IL algorithms frequently recover rewards which induce policies far weaker than the expert, even in the same environment the demonstrations were collected in. Furthermore, (b) these rewards are often quite poorly shaped, necessitating extensive environment interaction to optimise effectively. We provide simple and scalable fixes to both of these concerns. For (a), we find that reward model ensembles combined with a slightly different training objective significantly improves re-training and transfer performance. For (b), we propose a novel evolution-strategies based method EvIL to optimise for a reward-shaping term that speeds up re-training in the target environment, closing a gap left open by the classical theory of IRL. On a suite of continuous control tasks, we are able to re-train policies in target (and source) environments more interaction-efficiently than prior work.
We improve the accuracy of Guidance & Control Networks (G&CNETs), trained to represent the optimal control policies of a time-optimal transfer and a mass-optimal landing, respectively. In both cases we leverage the dynamics of the spacecraft, described by Ordinary Differential Equations which incorporate a neural network on their right-hand side (Neural ODEs). Since the neural dynamics is differentiable, the ODEs sensitivities to the network parameters can be computed using the variational equations, thereby allowing to update the G&CNET parameters based on the observed dynamics. We start with a straightforward regression task, training the G&CNETs on datasets of optimal trajectories using behavioural cloning. These networks are then refined using the Neural ODE sensitivities by minimizing the error between the final states and the target states. We demonstrate that for the orbital transfer, the final error to the target can be reduced by 99% on a single trajectory and by 70% on a batch of 500 trajectories. For the landing problem the reduction in error is around 98-99% (position) and 40-44% (velocity). This step significantly enhances the accuracy of G&CNETs, which instills greater confidence in their reliability for operational use. We also compare our results to the popular Dataset Aggregation method (DaGGER) and allude to the strengths and weaknesses of both methods.
We introduce Language Feedback Models (LFMs) that identify desirable behaviour - actions that help achieve tasks specified in the instruction - for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFM can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.
Anomaly detection in decision-making sequences is a challenging problem due to the complexity of normality representation learning and the sequential nature of the task. Most existing methods based on Reinforcement Learning (RL) are difficult to implement in the real world due to unrealistic assumptions, such as having access to environment dynamics, reward signals, and online interactions with the environment. To address these limitations, we propose an unsupervised method named Offline Imitation Learning based Anomaly Detection (OIL-AD), which detects anomalies in decision-making sequences using two extracted behaviour features: action optimality and sequential association. Our offline learning model is an adaptation of behavioural cloning with a transformer policy network, where we modify the training process to learn a Q function and a state value function from normal trajectories. We propose that the Q function and the state value function can provide sufficient information about agents' behavioural data, from which we derive two features for anomaly detection. The intuition behind our method is that the action optimality feature derived from the Q function can differentiate the optimal action from others at each local state, and the sequential association feature derived from the state value function has the potential to maintain the temporal correlations between decisions (state-action pairs). Our experiments show that OIL-AD can achieve outstanding online anomaly detection performance with up to 34.8% improvement in F1 score over comparable baselines.
Game development is a long process that involves many stages before a product is ready for the market. Human play testing is among the most time consuming, as testers are required to repeatedly perform tasks in the search for errors in the code. Therefore, automated testing is seen as a key technology for the gaming industry, as it would dramatically improve development costs and efficiency. Toward this end, we propose EVOLUTE, a novel imitation learning-based architecture that combines behavioural cloning (BC) with energy based models (EBMs). EVOLUTE is a two-stream ensemble model that splits the action space of autonomous agents into continuous and discrete tasks. The EBM stream handles the continuous tasks, to have a more refined and adaptive control, while the BC stream handles discrete actions, to ease training. We evaluate the performance of EVOLUTE in a shooting-and-driving game, where the agent is required to navigate and continuously identify targets to attack. The proposed model has higher generalisation capabilities than standard BC approaches, showing a wider range of behaviours and higher performances. Also, EVOLUTE is easier to train than a pure end-to-end EBM model, as discrete tasks can be quite sparse in the dataset and cause model training to explore a much wider set of possible actions while training.
This paper describes methods for training autonomous agents to play the game "Doom 2" through Imitation Learning (IL) using only pixel data as input. We also explore how Reinforcement Learning (RL) compares to IL for humanness by comparing camera movement and trajectory data. Through behavioural cloning, we examine the ability of individual models to learn varying behavioural traits. We attempt to mimic the behaviour of real players with different play styles, and find we can train agents that behave aggressively, passively, or simply more human-like than traditional AIs. We propose these methods of introducing more depth and human-like behaviour to agents in video games. The trained IL agents perform on par with the average players in our dataset, whilst outperforming the worst players. While performance was not as strong as common RL approaches, it provides much stronger human-like behavioural traits to the agent.
Imitation learning with visual observations is notoriously inefficient when addressed with end-to-end behavioural cloning methods. In this paper, we explore an alternative paradigm which decomposes reasoning into three phases. First, a retrieval phase, which informs the robot what it can do with an object. Second, an alignment phase, which informs the robot where to interact with the object. And third, a replay phase, which informs the robot how to interact with the object. Through a series of real-world experiments on everyday tasks, such as grasping, pouring, and inserting objects, we show that this decomposition brings unprecedented learning efficiency, and effective inter- and intra-class generalisation. Videos are available at https://www.robot-learning.uk/retrieval-alignment-replay.
We present placing via picking (PvP), a method to autonomously collect real-world demonstrations for a family of placing tasks in which objects must be manipulated to specific, contact-constrained locations. With PvP, we approach the collection of robotic object placement demonstrations by reversing the grasping process and exploiting the inherent symmetry of the pick and place problems. Specifically, we obtain placing demonstrations from a set of grasp sequences of objects initially located at their target placement locations. Our system can collect hundreds of demonstrations in contact-constrained environments without human intervention using two modules: compliant control for grasping and tactile regrasping. We train a policy directly from visual observations through behavioural cloning, using the autonomously-collected demonstrations. By doing so, the policy can generalize to object placement scenarios outside of the training environment without privileged information (e.g., placing a plate picked up from a table). We validate our approach in home robot scenarios that include dishwasher loading and table setting. Our approach yields robotic placing policies that outperform policies trained with kinesthetic teaching, both in terms of success rate and data efficiency, while requiring no human supervision.
Solving long sequential tasks poses a significant challenge in embodied artificial intelligence. Enabling a robotic system to perform diverse sequential tasks with a broad range of manipulation skills is an active area of research. In this work, we present a Hybrid Hierarchical Learning framework, the Robotic Manipulation Network (ROMAN), to address the challenge of solving multiple complex tasks over long time horizons in robotic manipulation. ROMAN achieves task versatility and robust failure recovery by integrating behavioural cloning, imitation learning, and reinforcement learning. It consists of a central manipulation network that coordinates an ensemble of various neural networks, each specialising in distinct re-combinable sub-tasks to generate their correct in-sequence actions for solving complex long-horizon manipulation tasks. Experimental results show that by orchestrating and activating these specialised manipulation experts, ROMAN generates correct sequential activations for accomplishing long sequences of sophisticated manipulation tasks and achieving adaptive behaviours beyond demonstrations, while exhibiting robustness to various sensory noises. These results demonstrate the significance and versatility of ROMAN's dynamic adaptability featuring autonomous failure recovery capabilities, and highlight its potential for various autonomous manipulation tasks that demand adaptive motor skills.
Behavioural cloning uses a dataset of demonstrations to learn a behavioural policy. To overcome various learning and policy adaptation problems, we propose to use latent space to index a demonstration dataset, instantly access similar relevant experiences, and copy behavior from these situations. Actions from a selected similar situation can be performed by the agent until representations of the agent's current situation and the selected experience diverge in the latent space. Thus, we formulate our control problem as a search problem over a dataset of experts' demonstrations. We test our approach on BASALT MineRL-dataset in the latent representation of a Video PreTraining model. We compare our model to state-of-the-art Minecraft agents. Our approach can effectively recover meaningful demonstrations and show human-like behavior of an agent in the Minecraft environment in a wide variety of scenarios. Experimental results reveal that performance of our search-based approach is comparable to trained models, while allowing zero-shot task adaptation by changing the demonstration examples.
Behavioural cloning is an imitation learning technique that teaches an agent how to behave via expert demonstrations. Recent approaches use self-supervision of fully-observable unlabelled snapshots of the states to decode state pairs into actions. However, the iterative learning scheme employed by these techniques is prone to get trapped into bad local minima. Previous work uses goal-aware strategies to solve this issue. However, this requires manual intervention to verify whether an agent has reached its goal. We address this limitation by incorporating a discriminator into the original framework, offering two key advantages and directly solving a learning problem previous work had. First, it disposes of the manual intervention requirement. Second, it helps in learning by guiding function approximation based on the state transition of the expert's trajectories. Third, the discriminator solves a learning issue commonly present in the policy model, which is to sometimes perform a `no action' within the environment until the agent finally halts.