Jeremy Wyatt

Robert Gieselmann, Mihai Samson, Federico Pecora et al.

Generative models have emerged as a powerful paradigm for AI planning, yet their performance remains constrained by the training data distribution. One approach is to improve generated solutions during inference by scaling test-time compute. A more efficient alternative is to optimize the inference process itself. In this paper, we show that a modified version of a classical Open-Closed List (OCL) search provides just such an efficient inference procedure. Our algorithm synergizes two learned components: a generative model that performs fast rollouts from intermediate states and a heuristic model that prioritizes among candidate reasoning paths. Key contributions include novel exploration control mechanisms and integration of learned models within the OCL framework. Across multiple combinatorial planning domains, our approach outperforms both neurosymbolic search baselines and classical solvers in computational efficiency and solution quality.

Robert Gieselmann, Henrike von Huelsen, Mihai Samson et al.

Generative models trained on synthetic plan data are a promising approach to generalized planning. Recent work has focused on finding any valid plan, rather than a high-quality solution. We address the challenge of producing high-quality plans, a computationally hard problem, in sub-exponential time. First, we demonstrate that, given optimal data, a decoder-only transformer can generate high-quality plans for unseen problem instances. Second, we show how to self-improve an initial model trained on sub-optimal data. Each round of self-improvement combines multiple model calls with graph search to generate improved plans, used for model fine-tuning. An experimental study on four domains: Blocksworld, Logistics, Labyrinth, and Sokoban, shows on average a 30% reduction in plan length over the source symbolic planner, with over 80% of plans being optimal, where the optimum is known. Plan quality is further improved by inference-time search. The model's latency scales sub-exponentially in contrast to the satisficing and optimal symbolic planners to which we compare. Together, these results suggest that self-improvement with generative models offers a scalable approach for high-quality plan generation.

A. Barsky, C. Zito, H. Mori et al.

The hype about sensorimotor learning is currently reaching high fever, thanks to the latest advancement in deep learning. In this paper, we present an open-source framework for collecting large-scale, time-synchronised synthetic data from highly disparate sensory modalities, such as audio, video, and proprioception, for learning robot manipulation tasks. We demonstrate the learning of non-linear sensorimotor mappings for a humanoid drumming robot that generates novel motion sequences from desired audio data using cross-modal correspondences. We evaluate our system through the quality of its cross-modal retrieval, for generating suitable motion sequences to match desired unseen audio or video sequences.

Umit Rusen Aktas, Chao Zhao, Marek Kopicki et al.

Dexterous grasping of a novel object given a single view is an open problem. This paper makes several contributions to its solution. First, we present a simulator for generating and testing dexterous grasps. Second we present a data set, generated by this simulator, of 2.4 million simulated dexterous grasps of variations of 294 base objects drawn from 20 categories. Third, we present a basic architecture for generation and evaluation of dexterous grasps that may be trained in a supervised manner. Fourth, we present three different evaluative architectures, employing ResNet-50 or VGG16 as their visual backbone. Fifth, we train, and evaluate seventeen variants of generative-evaluative architectures on this simulated data set, showing improvement from 69.53% grasp success rate to 90.49%. Finally, we present a real robot implementation and evaluate the four most promising variants, executing 196 real robot grasps in total. We show that our best architectural variant achieves a grasp success rate of 87.8% on real novel objects seen from a single view, improving on a baseline of 57.1%.

Marek Kopicki, Dominik Belter, Jeremy L. Wyatt

This paper concerns the problem of how to learn to grasp dexterously, so as to be able to then grasp novel objects seen only from a single view-point. Recently, progress has been made in data-efficient learning of generative grasp models which transfer well to novel objects. These generative grasp models are learned from demonstration (LfD). One weakness is that, as this paper shall show, grasp transfer under challenging single view conditions is unreliable. Second, the number of generative model elements rises linearly in the number of training examples. This, in turn, limits the potential of these generative models for generalisation and continual improvement. In this paper, it is shown how to address these problems. Several technical contributions are made: (i) a view-based model of a grasp; (ii) a method for combining and compressing multiple grasp models; (iii) a new way of evaluating contacts that is used both to generate and to score grasps. These, together, improve both grasp performance and reduce the number of models learned for grasp transfer. These advances, in turn, also allow the introduction of autonomous training, in which the robot learns from self-generated grasps. Evaluation on a challenging test set shows that, with innovations (i)-(iii) deployed, grasp transfer success rises from 55.1% to 81.6%. By adding autonomous training this rises to 87.8%. These differences are statistically significant. In total, across all experiments, 539 test grasps were executed on real objects.

Ermano Arruda, Claudio Zito, Mohan Sridharan et al.

We present a parametric formulation for learning generative models for grasp synthesis from a demonstration. We cast new light on this family of approaches, proposing a parametric formulation for grasp synthesis that is computationally faster compared to related work and indicates better grasp success rate performance in simulated experiments, showing a gain of at least 10% success rate (p < 0.05) in all the tested conditions. The proposed implementation is also able to incorporate arbitrary constraints for grasp ranking that may include task-specific constraints. Results are reported followed by a brief discussion on the merits of the proposed methods noted so far.

Claudio Zito, Valerio Ortenzi, Maxime Adjigble et al.

Belief space planning is a viable alternative to formalise partially observable control problems and, in the recent years, its application to robot manipulation problems has grown. However, this planning approach was tried successfully only on simplified control problems. In this paper, we apply belief space planning to the problem of planning dexterous reach-to-grasp trajectories under object pose uncertainty. In our framework, the robot perceives the object to be grasped on-the-fly as a point cloud and compute a full 6D, non-Gaussian distribution over the object's pose (our belief space). The system has no limitations on the geometry of the object, i.e., non-convex objects can be represented, nor assumes that the point cloud is a complete representation of the object. A plan in the belief space is then created to reach and grasp the object, such that the information value of expected contacts along the trajectory is maximised to compensate for the pose uncertainty. If an unexpected contact occurs when performing the action, such information is used to refine the pose distribution and triggers a re-planning. Experimental results show that our planner (IR3ne) improves grasp reliability and compensates for the pose uncertainty such that it doubles the proportion of grasps that succeed on a first attempt.

Ermano Arruda, Michael J Mathew, Marek Kopicki et al.

Planning robust robot manipulation requires good forward models that enable robust plans to be found. This work shows how to achieve this using a forward model learned from robot data to plan push manipulations. We explore learning methods (Gaussian Process Regression, and an Ensemble of Mixture Density Networks) that give estimates of the uncertainty in their predictions. These learned models are utilised by a model predictive path integral (MPPI) controller to plan how to push the box to a goal location. The planner avoids regions of high predictive uncertainty in the forward model. This includes both inherent uncertainty in dynamics, and meta uncertainty due to limited data. Thus, pushing tasks are completed in a robust fashion with respect to estimated uncertainty in the forward model and without the need of differentiable cost functions. We demonstrate the method on a real robot, and show that learning can outperform physics simulation. Using simulation, we also show the ability to plan uncertainty averse paths.

Ermano Arruda, Jeremy Wyatt, Marek Kopicki

How should a robot direct active vision so as to ensure reliable grasping? We answer this question for the case of dexterous grasping of unfamiliar objects. By dexterous grasping we simply mean grasping by any hand with more than two fingers, such that the robot has some choice about where to place each finger. Such grasps typically fail in one of two ways, either unmodeled objects in the scene cause collisions or object reconstruction is insufficient to ensure that the grasp points provide a stable force closure. These problems can be solved more easily if active sensing is guided by the anticipated actions. Our approach has three stages. First, we take a single view and generate candidate grasps from the resulting partial object reconstruction. Second, we drive the active vision approach to maximise surface reconstruction quality around the planned contact points. During this phase, the anticipated grasp is continually refined. Third, we direct gaze to improve the safety of the planned reach to grasp trajectory. We show, on a dexterous manipulator with a camera on the wrist, that our approach (80.4% success rate) outperforms a randomised algorithm (64.3% success rate).

Marek Kopicki, Carlos J. Rosales, Hamal Marino et al.

Recent advances have been made in learning of grasps for fully actuated hands. A typical approach learns the target locations of finger links on the object. When a new object must be grasped, new finger locations are generated, and a collision free reach-to-grasp trajectory is planned. This assumes a collision free trajectory to the final grasp. This is not possible with underactuated hands, which cannot be guaranteed to avoid contact, and in fact exploit contacts with the object during grasping, so as to reach an equilibrium state in which the object is held securely. Unfortunately, these contact interactions are i) not directly controllable, and ii) hard to monitor during a real grasp. We overcome these problems so as to permit learning of transferrable grasps for underactuated hands. We make two main technical innovations. First, we model contact interactions during the grasp implicitly. We do this by modelling motor commands that lead reliably to the equilibrium state, rather than modelling contact changes themselves. This alters our reach-to-grasp model. Second, we extend our contact model learning algorithm to work with multiple training examples for each grasp type. This requires the ability to learn which parts of the hand reliably interact with the object during a particular grasp. Our approach learns from a rigid body simulation. This enables us to learn how to approach the object and close the underactuated hand from a variety of poses. From nine training grasps on three objects the method transferred grasps to previously unseen, novel objects, that differ significantly from the training objects, with an 80% success rate.

Domen Tabernik, Matej Kristan, Jeremy L. Wyatt et al.

Hierarchical feature learning based on convolutional neural networks (CNN) has recently shown significant potential in various computer vision tasks. While allowing high-quality discriminative feature learning, the downside of CNNs is the lack of explicit structure in features, which often leads to overfitting, absence of reconstruction from partial observations and limited generative abilities. Explicit structure is inherent in hierarchical compositional models, however, these lack the ability to optimize a well-defined cost function. We propose a novel analytic model of a basic unit in a layered hierarchical model with both explicit compositional structure and a well-defined discriminative cost function. Our experiments on two datasets show that the proposed compositional model performs on a par with standard CNNs on discriminative tasks, while, due to explicit modeling of the structure in the feature units, affording a straight-forward visualization of parts and faster inference due to separability of the units. Actions

Nick Hawes, Chris Burbridge, Ferdian Jovan et al.

Thanks to the efforts of the robotics and autonomous systems community, robots are becoming ever more capable. There is also an increasing demand from end-users for autonomous service robots that can operate in real environments for extended periods. In the STRANDS project we are tackling this demand head-on by integrating state-of-the-art artificial intelligence and robotics research into mobile service robots, and deploying these systems for long-term installations in security and care environments. Over four deployments, our robots have been operational for a combined duration of 104 days autonomously performing end-user defined tasks, covering 116km in the process. In this article we describe the approach we have used to enable long-term autonomous operation in everyday environments, and how our robots are able to use their long run times to improve their own performance.

Mohan Sridharan, Michael Gelfond, Shiqi Zhang et al.

This paper describes an architecture for robots that combines the complementary strengths of probabilistic graphical models and declarative programming to represent and reason with logic-based and probabilistic descriptions of uncertainty and domain knowledge. An action language is extended to support non-boolean fluents and non-deterministic causal laws. This action language is used to describe tightly-coupled transition diagrams at two levels of granularity, with a fine-resolution transition diagram defined as a refinement of a coarse-resolution transition diagram of the domain. The coarse-resolution system description, and a history that includes (prioritized) defaults, are translated into an Answer Set Prolog (ASP) program. For any given goal, inference in the ASP program provides a plan of abstract actions. To implement each such abstract action, the robot automatically zooms to the part of the fine-resolution transition diagram relevant to this action. A probabilistic representation of the uncertainty in sensing and actuation is then included in this zoomed fine-resolution system description, and used to construct a partially observable Markov decision process (POMDP). The policy obtained by solving the POMDP is invoked repeatedly to implement the abstract action as a sequence of concrete actions, with the corresponding observations being recorded in the coarse-resolution history and used for subsequent reasoning. The architecture is evaluated in simulation and on a mobile robot moving objects in an indoor domain, to show that it supports reasoning with violation of defaults, noisy observations and unreliable actions, in complex domains.

Shiqi Zhang, Mohan Sridharan, Michael Gelfond et al.

This paper describes an architecture that combines the complementary strengths of declarative programming and probabilistic graphical models to enable robots to represent, reason with, and learn from, qualitative and quantitative descriptions of uncertainty and knowledge. An action language is used for the low-level (LL) and high-level (HL) system descriptions in the architecture, and the definition of recorded histories in the HL is expanded to allow prioritized defaults. For any given goal, tentative plans created in the HL using default knowledge and commonsense reasoning are implemented in the LL using probabilistic algorithms, with the corresponding observations used to update the HL history. Tight coupling between the two levels enables automatic selection of relevant variables and generation of suitable action policies in the LL for each HL action, and supports reasoning with violation of defaults, noisy observations and unreliable actions in large and complex domains. The architecture is evaluated in simulation and on physical robots transporting objects in indoor domains; the benefit on robots is a reduction in task execution time of 39% compared with a purely probabilistic, but still hierarchical, approach.