Traversing the Reality Gap via Simulator Tuning

Collins, Jack; Brown, Ross; Leitner, Jurgen; Howard, David

doi:10.48550/arxiv.2003.01369

Cited by 5 publications

(13 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Simulator tuning. The concern of the reality gap in simulators has been highlighted in the past, albeit in domains other than resource management 13,18 . The root cause of simulations being far from realism is the improper tuning of the simulation parameters for diverse scenarios that need to be simulated.…”

mentioning

confidence: 99%

“…Considering the growing complexity of modern simulators with millions of parameters, each with millions of possible values, running a brute-force approach is intractable. To address this, prior work has leveraged evolutionary optimization strategies such as Sim2Real 18 . Sim2Real iteratively updates the simulator parameters, performs simulations and evaluates the deviation between the simulated and true values.…”

mentioning

confidence: 99%

See 1 more Smart Citation

SimTune: bridging the simulator reality gap for resource management in edge-cloud computing

Tuli

Casale

Jennings

2022

Sci Rep

View full text Add to dashboard Cite

Industries and services are undergoing an Internet of Things centric transformation globally, giving rise to an explosion of multi-modal data generated each second. This, with the requirement of low-latency result delivery, has led to the ubiquitous adoption of edge and cloud computing paradigms. Edge computing follows the data gravity principle, wherein the computational devices move closer to the end-users to minimize data transfer and communication times. However, large-scale computation has exacerbated the problem of efficient resource management in hybrid edge-cloud platforms. In this regard, data-driven models such as deep neural networks (DNNs) have gained popularity to give rise to the notion of edge intelligence. However, DNNs face significant problems of data saturation when fed volatile data. Data saturation is when providing more data does not translate to improvements in performance. To address this issue, prior work has leveraged coupled simulators that, akin to digital twins, generate out-of-distribution training data alleviating the data-saturation problem. However, simulators face the reality-gap problem, which is the inaccuracy in the emulation of real computational infrastructure due to the abstractions in such simulators. To combat this, we develop a framework, SimTune, that tackles this challenge by leveraging a low-fidelity surrogate model of the high-fidelity simulator to update the parameters of the latter, so to increase the simulation accuracy. This further helps co-simulated methods to generalize to edge-cloud configurations for which human encoded parameters are not known apriori. Experiments comparing SimTune against state-of-the-art data-driven resource management solutions on a real edge-cloud platform demonstrate that simulator tuning can improve quality of service metrics such as energy consumption and response time by up to 14.7% and 7.6% respectively.

show abstract

mentioning

confidence: 99%

mentioning

confidence: 99%

SimTune: bridging the simulator reality gap for resource management in edge-cloud computing

Tuli

Casale

Jennings

2022

Sci Rep

View full text Add to dashboard Cite

show abstract

“…An interesting question that arises from the different pretrainings is whether the neural nets trained with only simulated images are good enough models for classifying real photographs. The difference between the performance of a model trained with real images and the same model trained with simulations is usually called the reality gap [37,10].…”

Section: Reality Gapmentioning

confidence: 99%

“…The use of synthetic data is very appealing in other areas as well, such as for deep reinforcement learning [30]. Multiple synthetic datasets [28,19,40] or engines [15,11,10] to generate them are now available for AI research. Until recently the domain gap between the synthetic dataset and the real one usually made synthetic-only training non-competitive, but with today's rendering programs the generalization error is becoming comparable to that between two similar real-life datasets [25].…”

Section: Introductionmentioning

confidence: 99%

Learning from Scarce Information: Using Synthetic Data to Classify Roman Fine Ware Pottery

Jareño

Helden

Mirkes

et al. 2021

Entropy

View full text Add to dashboard Cite

In this article, we consider a version of the challenging problem of learning from datasets whose size is too limited to allow generalisation beyond the training set. To address the challenge, we propose to use a transfer learning approach whereby the model is first trained on a synthetic dataset replicating features of the original objects. In this study, the objects were smartphone photographs of near-complete Roman terra sigillata pottery vessels from the collection of the Museum of London. Taking the replicated features from published profile drawings of pottery forms allowed the integration of expert knowledge into the process through our synthetic data generator. After this first initial training the model was fine-tuned with data from photographs of real vessels. We show, through exhaustive experiments across several popular deep learning architectures, different test priors, and considering the impact of the photograph viewpoint and excessive damage to the vessels, that the proposed hybrid approach enables the creation of classifiers with appropriate generalisation performance. This performance is significantly better than that of classifiers trained exclusively on the original data, which shows the promise of the approach to alleviate the fundamental issue of learning from small datasets.

show abstract

“…The inconsistency between physical parameters and dynamics such as gravity, friction, collisions, density and simulating soft surfaces are complex to model in a virtual environment. Also, noise introduced by sensors, gear backslash, latency and environment factors such as texture quality, reflection and refraction could adversely affect performance of the trained model [3]. This phenomenon is known as the Reality Gap.…”

Section: Introductionmentioning

confidence: 99%

Bridging Reality Gap Between Virtual and Physical Robot through Domain Randomization and Induced Noise

Ranaweera¹,

Mahmoud²

2022

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

This paper investigates techniques that can be utilized to bridge the reality gap between virtual and physical robots, by implementing a virtual environment and a physical robotic platform to evaluate the robustness of transfer learning from virtual to real-world robots. The proposed approach utilizes two reinforcement (RL) learning methods: deep Q-learning and Actor-Critic methodology to create a model that can learn from a virtual environment and performs in a physical environment. Techniques such as domain randomization and induced noise during training were utilized to bring variability and ultimately improve the learning policies. The experimental results demonstrate the effectiveness of the Actor-Critic reinforcement learning technique to bridge the reality gap.

show abstract

Traversing the Reality Gap via Simulator Tuning

Cited by 5 publications

References 26 publications

SimTune: bridging the simulator reality gap for resource management in edge-cloud computing

SimTune: bridging the simulator reality gap for resource management in edge-cloud computing

Learning from Scarce Information: Using Synthetic Data to Classify Roman Fine Ware Pottery

Bridging Reality Gap Between Virtual and Physical Robot through Domain Randomization and Induced Noise

Contact Info

Product

Resources

About