Jürgen Leitner

Yusuke Miyashita, Dimitris Gahtidis, Colin La et al.

In this paper, we propose the use of generative artificial intelligence (AI) to improve zero-shot performance of a pre-trained policy by altering observations during inference. Modern robotic systems, powered by advanced neural networks, have demonstrated remarkable capabilities on pre-trained tasks. However, generalizing and adapting to new objects and environments is challenging, and fine-tuning visuomotor policies is time-consuming. To overcome these issues we propose Robotic Policy Inference via Synthetic Observations (ROSO). ROSO uses stable diffusion to pre-process a robot's observation of novel objects during inference time to fit within its distribution of observations of the pre-trained policies. This novel paradigm allows us to transfer learned knowledge from known tasks to previously unseen scenarios, enhancing the robot's adaptability without requiring lengthy fine-tuning. Our experiments show that incorporating generative AI into robotic inference significantly improves successful outcomes, finishing up to 57% of tasks otherwise unsuccessful with the pre-trained policy.

Vibhavari Dasagi, Robert Lee, Jake Bruce et al.

Deep reinforcement learning has been shown to solve challenging tasks where large amounts of training experience is available, usually obtained online while learning the task. Robotics is a significant potential application domain for many of these algorithms, but generating robot experience in the real world is expensive, especially when each task requires a lengthy online training procedure. Off-policy algorithms can in principle learn arbitrary tasks from a diverse enough fixed dataset. In this work, we evaluate popular exploration methods by generating robotics datasets for the purpose of learning to solve tasks completely offline without any further interaction in the real world. We present results on three popular continuous control tasks in simulation, as well as continuous control of a high-dimensional real robot arm. Code documenting all algorithms, experiments, and hyper-parameters is available at https://github.com/qutrobotlearning/batchlearning.

S. Wade-McCue, N. Kelly-Boxall, M. McTaggart et al.

We present the grasping system and design approach behind Cartman, the winning entrant in the 2017 Amazon Robotics Challenge. We investigate the design processes leading up to the final iteration of the system and describe the emergent solution by comparing it with key robotics design aspects. Following our experience, we propose a new design aspect, precision vs. redundancy, that should be considered alongside the previously proposed design aspects of modularity vs. integration, generality vs. assumptions, computation vs. embodiment and planning vs. feedback. We present the grasping system behind Cartman, the winning robot in the 2017 Amazon Robotics Challenge. The system makes strong use of redundancy in design by implementing complimentary tools, a suction gripper and a parallel gripper. This multi-modal end-effector is combined with three grasp synthesis algorithms to accommodate the range of objects provided by Amazon during the challenge. We provide a detailed system description and an evaluation of its performance before discussing the broader nature of the system with respect to the key aspects of robotic design as initially proposed by the winners of the first Amazon Picking Challenge. To address the principal nature of our grasping system and the reason for its success, we propose an additional robotic design aspect `precision vs. redundancy'. The full design of our robotic system, including the end-effector, is open sourced and available at http://juxi.net/projects/AmazonRoboticsChallenge/

M. McTaggart, D. Morrison, A. W. Tow et al.

Robotic manipulation and grasping in cluttered and unstructured environments is a current challenge for robotics. Enabling robots to operate in these challenging environments have direct applications from automating warehouses to harvesting fruit in agriculture. One of the main challenges associated with these difficult robotic manipulation tasks is the motion planning and control problem for multi-DoF (Degree of Freedom) manipulators. This paper presents the design and performance evaluation of a low-cost Cartesian manipulator, Cartman who took first place in the Amazon Robotics Challenge 2017. It can perform pick and place tasks of household items in a cluttered environment. The robot is capable of linear speeds of 1 m/s and angular speeds of 1.5 rad/s, capable of sub-millimetre static accuracy and safe payload capacity of 2kg. Cartman can be produced for under 10 000 AUD. The complete design is open sourced and can be found at http://juxi.net/projects/AmazonRoboticsChallenge.

Ming Hu, Peng Xia, Lin Wang et al.

Surgical scene perception via videos is critical for advancing robotic surgery, telesurgery, and AI-assisted surgery, particularly in ophthalmology. However, the scarcity of diverse and richly annotated video datasets has hindered the development of intelligent systems for surgical workflow analysis. Existing datasets face challenges such as small scale, lack of diversity in surgery and phase categories, and absence of time-localized annotations. These limitations impede action understanding and model generalization validation in complex and diverse real-world surgical scenarios. To address this gap, we introduce OphNet, a large-scale, expert-annotated video benchmark for ophthalmic surgical workflow understanding. OphNet features: 1) A diverse collection of 2,278 surgical videos spanning 66 types of cataract, glaucoma, and corneal surgeries, with detailed annotations for 102 unique surgical phases and 150 fine-grained operations. 2) Sequential and hierarchical annotations for each surgery, phase, and operation, enabling comprehensive understanding and improved interpretability. 3) Time-localized annotations, facilitating temporal localization and prediction tasks within surgical workflows. With approximately 285 hours of surgical videos, OphNet is about 20 times larger than the largest existing surgical workflow analysis benchmark. Code and dataset are available at: https://minghu0830.github.io/OphNet-benchmark/.

Jack Collins, Ross Brown, Jürgen Leitner et al.

We propose a novel iterative approach for crossing the reality gap that utilises live robot rollouts and differentiable physics. Our method, RealityGrad, demonstrates for the first time, an efficient sim2real transfer in combination with a real2sim model optimisation for closing the reality gap. Differentiable physics has become an alluring alternative to classical rigid-body simulation due to the current culmination of automatic differentiation libraries, compute and non-linear optimisation libraries. Our method builds on this progress and employs differentiable physics for efficient trajectory optimisation. We demonstrate RealitGrad on a dynamic control task for a serial link robot manipulator and present results that show its efficiency and ability to quickly improve not just the robot's performance in real world tasks but also enhance the simulation model for future tasks. One iteration of RealityGrad takes less than 22 minutes on a desktop computer while reducing the error by 2/3, making it efficient compared to other sim2real methods in both compute and time. Our methodology and application of differentiable physics establishes a promising approach for crossing the reality gap and has great potential for scaling to complex environments.

Brendan Tidd, Akansel Cosgun, Jurgen Leitner et al.

The U.S. Defense Advanced Research Projects Agency (DARPA) Subterranean Challenge requires teams of robots to traverse difficult and diverse underground environments. Traversing small gaps is one of the challenging scenarios that robots encounter. Imperfect sensor information makes it difficult for classical navigation methods, where behaviours require significant manual fine tuning. In this paper we present a deep reinforcement learning method for autonomously navigating through small gaps, where contact between the robot and the gap may be required. We first learn a gap behaviour policy to get through small gaps (only centimeters wider than the robot). We then learn a goal-conditioned behaviour selection policy that determines when to activate the gap behaviour policy. We train our policies in simulation and demonstrate their effectiveness with a large tracked robot in simulation and on the real platform. In simulation experiments, our approach achieves 93\% success rate when the gap behaviour is activated manually by an operator, and 63\% with autonomous activation using the behaviour selection policy. In real robot experiments, our approach achieves a success rate of 73\% with manual activation, and 40\% with autonomous behaviour selection. While we show the feasibility of our approach in simulation, the difference in performance between simulated and real world scenarios highlight the difficulty of direct sim-to-real transfer for deep reinforcement learning policies. In both the simulated and real world environments alternative methods were unable to traverse the gap.

Brendan Tidd, Nicolas Hudson, Akansel Cosgun et al.

Dynamic platforms that operate over many unique terrain conditions typically require many behaviours. To transition safely, there must be an overlap of states between adjacent controllers. We develop a novel method for training setup policies that bridge the trajectories between pre-trained Deep Reinforcement Learning (DRL) policies. We demonstrate our method with a simulated biped traversing a difficult jump terrain, where a single policy fails to learn the task, and switching between pre-trained policies without setup policies also fails. We perform an ablation of key components of our system, and show that our method outperforms others that learn transition policies. We demonstrate our method with several difficult and diverse terrain types, and show that we can use setup policies as part of a modular control suite to successfully traverse a sequence of complex terrains. We show that using setup policies improves the success rate for traversing a single difficult jump terrain (from 51.3% success rate with the best comparative method to 82.2%), and traversing a random sequence of difficult obstacles (from 1.9% without setup policies to 71.2%).

Brendan Tidd, Nicolas Hudson, Akansel Cosgun et al.

Legged robots often use separate control policiesthat are highly engineered for traversing difficult terrain suchas stairs, gaps, and steps, where switching between policies isonly possible when the robot is in a region that is commonto adjacent controllers. Deep Reinforcement Learning (DRL)is a promising alternative to hand-crafted control design,though typically requires the full set of test conditions to beknown before training. DRL policies can result in complex(often unrealistic) behaviours that have few or no overlappingregions between adjacent policies, making it difficult to switchbehaviours. In this work we develop multiple DRL policieswith Curriculum Learning (CL), each that can traverse asingle respective terrain condition, while ensuring an overlapbetween policies. We then train a network for each destinationpolicy that estimates the likelihood of successfully switchingfrom any other policy. We evaluate our switching methodon a previously unseen combination of terrain artifacts andshow that it performs better than heuristic methods. Whileour method is trained on individual terrain types, it performscomparably to a Deep Q Network trained on the full set ofterrain conditions. This approach allows the development ofseparate policies in constrained conditions with embedded priorknowledge about each behaviour, that is scalable to any numberof behaviours, and prepares DRL methods for applications inthe real world

Robert Lee, Daniel Ward, Akansel Cosgun et al.

Manipulating deformable objects, such as fabric, is a long standing problem in robotics, with state estimation and control posing a significant challenge for traditional methods. In this paper, we show that it is possible to learn fabric folding skills in only an hour of self-supervised real robot experience, without human supervision or simulation. Our approach relies on fully convolutional networks and the manipulation of visual inputs to exploit learned features, allowing us to create an expressive goal-conditioned pick and place policy that can be trained efficiently with real world robot data only. Folding skills are learned with only a sparse reward function and thus do not require reward function engineering, merely an image of the goal configuration. We demonstrate our method on a set of towel-folding tasks, and show that our approach is able to discover sequential folding strategies, purely from trial-and-error. We achieve state-of-the-art results without the need for demonstrations or simulation, used in prior approaches. Videos available at: https://sites.google.com/view/learningtofold

Jack Collins, Ross Brown, Jurgen Leitner et al.

The large demand for simulated data has made the reality gap a problem on the forefront of robotics. We propose a method to traverse the gap by tuning available simulation parameters. Through the optimisation of physics engine parameters, we show that we are able to narrow the gap between simulated solutions and a real world dataset, and thus allow more ready transfer of leaned behaviours between the two. We subsequently gain understanding as to the importance of specific simulator parameters, which is of broad interest to the robotic machine learning community. We find that even optimised for different tasks that different physics engine perform better in certain scenarios and that friction and maximum actuator velocity are tightly bounded parameters that greatly impact the transference of simulated solutions.

Douglas Morrison, Peter Corke, Jürgen Leitner

We present the Evolved Grasping Analysis Dataset (EGAD), comprising over 2000 generated objects aimed at training and evaluating robotic visual grasp detection algorithms. The objects in EGAD are geometrically diverse, filling a space ranging from simple to complex shapes and from easy to difficult to grasp, compared to other datasets for robotic grasping, which may be limited in size or contain only a small number of object classes. Additionally, we specify a set of 49 diverse 3D-printable evaluation objects to encourage reproducible testing of robotic grasping systems across a range of complexity and difficulty. The dataset, code and videos can be found at https://dougsm.github.io/egad/

Jack Collins, Jessie McVicar, David Wedlock et al.

We present a benchmark to facilitate simulated manipulation; an attempt to overcome the obstacles of physical benchmarks through the distribution of a real world, ground truth dataset. Users are given various simulated manipulation tasks with assigned protocols having the objective of replicating the real world results of a recorded dataset. The benchmark comprises of a range of metrics used to characterise the successes of submitted environments whilst providing insight into their deficiencies. We apply our benchmark to two simulation environments, PyBullet and V-Rep, and publish the results. All materials required to benchmark an environment, including protocols and the dataset, can be found at the benchmarks' website https://research.csiro.au/robotics/manipulation-benchmark/.

Timo Korthals, Malte Schilling, Jürgen Leitner

This contribution comprises the interplay between a multi-modal variational autoencoder and an environment to a perceived environment, on which an agent can act. Furthermore, we conclude our work with a comparison to curiosity-driven learning.

Vibhavari Dasagi, Jake Bruce, Thierry Peynot et al.

When learning behavior, training data is often generated by the learner itself; this can result in unstable training dynamics, and this problem has particularly important applications in safety-sensitive real-world control tasks such as robotics. In this work, we propose a principled and model-agnostic approach to mitigate the issue of unstable learning dynamics by maintaining a history of a reinforcement learning agent over the course of training, and reverting to the parameters of a previous agent whenever performance significantly decreases. We develop techniques for evaluating this performance through statistical hypothesis testing of continued improvement, and evaluate them on a standard suite of challenging benchmark tasks involving continuous control of simulated robots. We show improvements over state-of-the-art reinforcement learning algorithms in performance and robustness to hyperparameters, outperforming DDPG in 5 out of 6 evaluation environments and showing no decrease in performance with TD3, which is known to be relatively stable. In this way, our approach takes an important step towards increasing data efficiency and stability in training for real-world robotic applications.

Jack Collins, David Howard, Jürgen Leitner

We quantify the accuracy of various simulators compared to a real world robotic reaching and interaction task. Simulators are used in robotics to design solutions for real world hardware without the need for physical access. The `reality gap' prevents solutions developed or learnt in simulation from performing well, or at at all, when transferred to real-world hardware. Making use of a Kinova robotic manipulator and a motion capture system, we record a ground truth enabling comparisons with various simulators, and present quantitative data for various manipulation-oriented robotic tasks. We show the relative strengths and weaknesses of numerous contemporary simulators, highlighting areas of significant discrepancy, and assisting researchers in the field in their selection of appropriate simulators for their use cases. All code and parameter listings are publicly available from: https://bitbucket.csiro.au/scm/~col549/quantifying-the-reality-gap-in-robotic-manipulation-tasks.git .

Douglas Morrison, Peter Corke, Jürgen Leitner

Camera viewpoint selection is an important aspect of visual grasp detection, especially in clutter where many occlusions are present. Where other approaches use a static camera position or fixed data collection routines, our Multi-View Picking (MVP) controller uses an active perception approach to choose informative viewpoints based directly on a distribution of grasp pose estimates in real time, reducing uncertainty in the grasp poses caused by clutter and occlusions. In trials of grasping 20 objects from clutter, our MVP controller achieves 80% grasp success, outperforming a single-viewpoint grasp detector by 12%. We also show that our approach is both more accurate and more efficient than approaches which consider multiple fixed viewpoints.

Vibhavari Dasagi, Robert Lee, Serena Mou et al.

Current end-to-end deep Reinforcement Learning (RL) approaches require jointly learning perception, decision-making and low-level control from very sparse reward signals and high-dimensional inputs, with little capability of incorporating prior knowledge. This results in prohibitively long training times for use on real-world robotic tasks. Existing algorithms capable of extracting task-level representations from high-dimensional inputs, e.g. object detection, often produce outputs of varying lengths, restricting their use in RL methods due to the need for neural networks to have fixed length inputs. In this work, we propose a framework that combines deep sets encoding, which allows for variable-length abstract representations, with modular RL that utilizes these representations, decoupling high-level decision making from low-level control. We successfully demonstrate our approach on the robot manipulation task of object sorting, showing that this method can learn effective policies within mere minutes of highly simplified simulation. The learned policies can be directly deployed on a robot without further training, and generalize to variations of the task unseen during training.

Timo Korthals, Jürgen Leitner, Ulrich Rückert

We investigate a reinforcement approach for distributed sensing based on the latent space derived from multi-modal deep generative models. Our contribution provides insights to the following benefits: Detections can be exchanged effectively between robots equipped with uni-modal sensors due to a shared latent representation of information that is trained by a Variational Auto Encoder (VAE). Sensor-fusion can be applied asynchronously due to the generative feature of the VAE. Deep Q-Networks (DQNs) are trained to minimize uncertainty in latent space by coordinating robots to a Point-of-Interest (PoI) where their sensor modality can provide beneficial information about the PoI. Additionally, we show that the decrease in uncertainty can be defined as the direct reward signal for training the DQN.

Niko Sünderhauf, Oliver Brock, Walter Scheirer et al.

The application of deep learning in robotics leads to very specific problems and research questions that are typically not addressed by the computer vision and machine learning communities. In this paper we discuss a number of robotics-specific learning, reasoning, and embodiment challenges for deep learning. We explain the need for better evaluation metrics, highlight the importance and unique challenges for deep robotic learning in simulation, and explore the spectrum between purely data-driven and model-driven approaches. We hope this paper provides a motivating overview of important research directions to overcome the current limitations, and help fulfill the promising potentials of deep learning in robotics.

Douglas Morrison, Peter Corke, Jürgen Leitner

This paper presents a real-time, object-independent grasp synthesis method which can be used for closed-loop grasping. Our proposed Generative Grasping Convolutional Neural Network (GG-CNN) predicts the quality and pose of grasps at every pixel. This one-to-one mapping from a depth image overcomes limitations of current deep-learning grasping techniques by avoiding discrete sampling of grasp candidates and long computation times. Additionally, our GG-CNN is orders of magnitude smaller while detecting stable grasps with equivalent performance to current state-of-the-art techniques. The light-weight and single-pass generative nature of our GG-CNN allows for closed-loop control at up to 50Hz, enabling accurate grasping in non-static environments where objects move and in the presence of robot control inaccuracies. In our real-world tests, we achieve an 83% grasp success rate on a set of previously unseen objects with adversarial geometry and 88% on a set of household objects that are moved during the grasp attempt. We also achieve 81% accuracy when grasping in dynamic clutter.

A. Milan, T. Pham, K. Vijay et al.

We present our approach for robotic perception in cluttered scenes that led to winning the recent Amazon Robotics Challenge (ARC) 2017. Next to small objects with shiny and transparent surfaces, the biggest challenge of the 2017 competition was the introduction of unseen categories. In contrast to traditional approaches which require large collections of annotated data and many hours of training, the task here was to obtain a robust perception pipeline with only few minutes of data acquisition and training time. To that end, we present two strategies that we explored. One is a deep metric learning approach that works in three separate steps: semantic-agnostic boundary detection, patch classification and pixel-wise voting. The other is a fully-supervised semantic segmentation approach with efficient dataset collection. We conduct an extensive analysis of the two methods on our ARC 2017 dataset. Interestingly, only few examples of each class are sufficient to fine-tune even very deep convolutional neural networks for this specific task.

D. Morrison, A. W. Tow, M. McTaggart et al.

The Amazon Robotics Challenge enlisted sixteen teams to each design a pick-and-place robot for autonomous warehousing, addressing development in robotic vision and manipulation. This paper presents the design of our custom-built, cost-effective, Cartesian robot system Cartman, which won first place in the competition finals by stowing 14 (out of 16) and picking all 9 items in 27 minutes, scoring a total of 272 points. We highlight our experience-centred design methodology and key aspects of our system that contributed to our competitiveness. We believe these aspects are crucial to building robust and effective robotic systems.

Fangyi Zhang, Jürgen Leitner, Zongyuan Ge et al.

Various approaches have been proposed to learn visuo-motor policies for real-world robotic applications. One solution is first learning in simulation then transferring to the real world. In the transfer, most existing approaches need real-world images with labels. However, the labelling process is often expensive or even impractical in many robotic applications. In this paper, we propose an adversarial discriminative sim-to-real transfer approach to reduce the cost of labelling real data. The effectiveness of the approach is demonstrated with modular networks in a table-top object reaching task where a 7 DoF arm is controlled in velocity mode to reach a blue cuboid in clutter through visual observations. The adversarial transfer approach reduced the labelled real data requirement by 50%. Policies can be transferred to real environments with only 93 labelled and 186 unlabelled real images. The transferred visuo-motor policies are robust to novel (not seen in training) objects in clutter and even a moving target, achieving a 97.8% success rate and 1.8 cm control accuracy.

Quentin Bateux, Eric Marchand, Jürgen Leitner et al.

We present a deep neural network-based method to perform high-precision, robust and real-time 6 DOF visual servoing. The paper describes how to create a dataset simulating various perturbations (occlusions and lighting conditions) from a single real-world image of the scene. A convolutional neural network is fine-tuned using this dataset to estimate the relative pose between two images of the same scene. The output of the network is then employed in a visual servoing control scheme. The method converges robustly even in difficult real-world settings with strong lighting variations and occlusions.A positioning error of less than one millimeter is obtained in experiments with a 6 DOF robot.

Fangyi Zhang, Jürgen Leitner, Michael Milford et al.

This paper introduces an end-to-end fine-tuning method to improve hand-eye coordination in modular deep visuo-motor policies (modular networks) where each module is trained independently. Benefiting from weighted losses, the fine-tuning method significantly improves the performance of the policies for a robotic planar reaching task.

Adam Tow, Niko Sünderhauf, Sareh Shirazi et al.

We propose to learn tasks directly from visual demonstrations by learning to predict the outcome of human and robot actions on an environment. We enable a robot to physically perform a human demonstrated task without knowledge of the thought processes or actions of the human, only their visually observable state transitions. We evaluate our approach on two table-top, object manipulation tasks and demonstrate generalisation to previously unseen states. Our approach reduces the priors required to implement a robot task learning system compared with the existing approaches of Learning from Demonstration, Reinforcement Learning and Inverse Reinforcement Learning.

Fangyi Zhang, Jürgen Leitner, Michael Milford et al.

While deep learning has had significant successes in computer vision thanks to the abundance of visual data, collecting sufficiently large real-world datasets for robot learning can be costly. To increase the practicality of these techniques on real robots, we propose a modular deep reinforcement learning method capable of transferring models trained in simulation to a real-world robotic task. We introduce a bottleneck between perception and control, enabling the networks to be trained independently, but then merged and fine-tuned in an end-to-end manner to further improve hand-eye coordination. On a canonical, planar visually-guided robot reaching task a fine-tuned accuracy of 1.6 pixels is achieved, a significant improvement over naive transfer (17.5 pixels), showing the potential for more complicated and broader applications. Our method provides a technique for more efficient learning and transfer of visuo-motor policies for real robotic systems without relying entirely on large real-world robot datasets.

Jürgen Leitner, Adam W. Tow, Jake E. Dean et al.

Robotic challenges like the Amazon Picking Challenge (APC) or the DARPA Challenges are an established and important way to drive scientific progress. They make research comparable on a well-defined benchmark with equal test conditions for all participants. However, such challenge events occur only occasionally, are limited to a small number of contestants, and the test conditions are very difficult to replicate after the main event. We present a new physical benchmark challenge for robotic picking: the ACRV Picking Benchmark (APB). Designed to be reproducible, it consists of a set of 42 common objects, a widely available shelf, and exact guidelines for object arrangement using stencils. A well-defined evaluation protocol enables the comparison of \emph{complete} robotic systems -- including perception and manipulation -- instead of sub-systems only. Our paper also describes and reports results achieved by an open baseline system based on a Baxter robot.

Donald G. Dansereau, Anders Eriksson, Jürgen Leitner

We generalize Richardson-Lucy (RL) deblurring to 4-D light fields by replacing the convolution steps with light field rendering of motion blur. The method deals correctly with blur caused by 6-degree-of-freedom camera motion in complex 3-D scenes, without performing depth estimation. We introduce a novel regularization term that maintains parallax information in the light field while reducing noise and ringing. We demonstrate the method operating effectively on rendered scenes and scenes captured using an off-the-shelf light field camera. An industrial robot arm provides repeatable and known trajectories, allowing us to establish quantitative performance in complex 3-D scenes. Qualitative and quantitative results confirm the effectiveness of the method, including commonly occurring cases for which previously published methods fail. We include mathematical proof that the algorithm converges to the maximum-likelihood estimate of the unblurred scene under Poisson noise. We expect extension to blind methods to be possible following the generalization of 2-D Richardson-Lucy to blind deconvolution.

Fangyi Zhang, Jürgen Leitner, Michael Milford et al.

This paper introduces a machine learning based system for controlling a robotic manipulator with visual perception only. The capability to autonomously learn robot controllers solely from raw-pixel images and without any prior knowledge of configuration is shown for the first time. We build upon the success of recent deep reinforcement learning and develop a system for learning target reaching with a three-joint robot manipulator using external visual observation. A Deep Q Network (DQN) was demonstrated to perform target reaching after training in simulation. Transferring the network to real hardware and real observation in a naive approach failed, but experiments show that the network works when replacing camera images with synthetic images.