Lars Ankile

Allen Z. Ren, Justin Lidard, Lars L. Ankile et al.

We introduce Diffusion Policy Policy Optimization, DPPO, an algorithmic framework including best practices for fine-tuning diffusion-based policies (e.g. Diffusion Policy) in continuous control and robot learning tasks using the policy gradient (PG) method from reinforcement learning (RL). PG methods are ubiquitous in training RL policies with other policy parameterizations; nevertheless, they had been conjectured to be less efficient for diffusion-based policies. Surprisingly, we show that DPPO achieves the strongest overall performance and efficiency for fine-tuning in common benchmarks compared to other RL methods for diffusion-based policies and also compared to PG fine-tuning of other policy parameterizations. Through experimental investigation, we find that DPPO takes advantage of unique synergies between RL fine-tuning and the diffusion parameterization, leading to structured and on-manifold exploration, stable training, and strong policy robustness. We further demonstrate the strengths of DPPO in a range of realistic settings, including simulated robotic tasks with pixel observations, and via zero-shot deployment of simulation-trained policies on robot hardware in a long-horizon, multi-stage manipulation task. Website with code: diffusion-ppo.github.io

L. L. Ankile, B. S. Ham, K. Mao et al.

When assisting human users in reinforcement learning (RL), we can represent users as RL agents and study key parameters, called \emph{user traits}, to inform intervention design. We study the relationship between user behaviors (policy classes) and user traits. Given an environment, we introduce an intuitive tool for studying the breakdown of "user types": broad sets of traits that result in the same behavior. We show that seemingly different real-world environments admit the same set of user types and formalize this observation as an equivalence relation defined on environments. By transferring intervention design between environments within the same equivalence class, we can help rapidly personalize interventions.

Lars Ankile, Anthony Simeonov, Idan Shenfeld et al.

Recent advances in Behavior Cloning (BC) have made it easy to teach robots new tasks. However, we find that the ease of teaching comes at the cost of unreliable performance that saturates with increasing data for tasks requiring precision. The performance saturation can be attributed to two critical factors: (a) distribution shift resulting from the use of offline data and (b) the lack of closed-loop corrective control caused by action chucking (predicting a set of future actions executed open-loop) critical for BC performance. Our key insight is that by predicting action chunks, BC policies function more like trajectory "planners" than closed-loop controllers necessary for reliable execution. To address these challenges, we devise a simple yet effective method, ResiP (Residual for Precise Manipulation), that overcomes the reliability problem while retaining BC's ease of teaching and long-horizon capabilities. ResiP augments a frozen, chunked BC model with a fully closed-loop residual policy trained with reinforcement learning (RL) that addresses distribution shifts and introduces closed-loop corrections over open-loop execution of action chunks predicted by the BC trajectory planner. Videos, code, and data: https://residual-assembly.github.io.

Lars Lien Ankile, Anna Midgley, Sebastian Weisshaar

Neural Networks are infamously sensitive to small perturbations in their inputs, making them vulnerable to adversarial attacks. This project evaluates the performance of Denoising Diffusion Probabilistic Models (DDPM) as a purification technique to defend against adversarial attacks. This works by adding noise to an adversarial example before removing it through the reverse process of the diffusion model. We evaluate the approach on the PatchCamelyon data set for histopathologic scans of lymph node sections and find an improvement of the robust accuracy by up to 88\% of the original model's accuracy, constituting a considerable improvement over the vanilla model and our baselines. The project code is located at https://github.com/ankile/Adversarial-Diffusion.

Lars Ankile, Anthony Simeonov, Idan Shenfeld et al.

While learning from demonstrations is powerful for acquiring visuomotor policies, high-performance imitation without large demonstration datasets remains challenging for tasks requiring precise, long-horizon manipulation. This paper proposes a pipeline for improving imitation learning performance with a small human demonstration budget. We apply our approach to assembly tasks that require precisely grasping, reorienting, and inserting multiple parts over long horizons and multiple task phases. Our pipeline combines expressive policy architectures and various techniques for dataset expansion and simulation-based data augmentation. These help expand dataset support and supervise the model with locally corrective actions near bottleneck regions requiring high precision. We demonstrate our pipeline on four furniture assembly tasks in simulation, enabling a manipulator to assemble up to five parts over nearly 2500 time steps directly from RGB images, outperforming imitation and data augmentation baselines. Project website: https://imitation-juicer.github.io/.

Marcel Torne, Arhan Jain, Jiayi Yuan et al.

Scaling robot learning requires data collection pipelines that scale favorably with human effort. In this work, we propose Crowdsourcing and Amortizing Human Effort for Real-to-Sim-to-Real(CASHER), a pipeline for scaling up data collection and learning in simulation where the performance scales superlinearly with human effort. The key idea is to crowdsource digital twins of real-world scenes using 3D reconstruction and collect large-scale data in simulation, rather than the real-world. Data collection in simulation is initially driven by RL, bootstrapped with human demonstrations. As the training of a generalist policy progresses across environments, its generalization capabilities can be used to replace human effort with model generated demonstrations. This results in a pipeline where behavioral data is collected in simulation with continually reducing human effort. We show that CASHER demonstrates zero-shot and few-shot scaling laws on three real-world tasks across diverse scenarios. We show that CASHER enables fine-tuning of pre-trained policies to a target scenario using a video scan without any additional human effort. See our project website: https://casher-robot-learning.github.io/CASHER/

Yash Yardi, Samuel Biruduganti, Lars Ankile

Simulation offers a scalable and efficient alternative to real-world data collection for learning visuomotor robotic policies. However, the simulation-to-reality, or Sim2Real distribution shift -- introduced by employing simulation-trained policies in real-world environments -- frequently prevents successful policy transfer. We present an offline framework to evaluate the performance of using large-scale pre-trained vision encoders to address the Sim2Real gap. We examine a diverse collection of encoders, assessing their ability to extract features necessary for robot control (Action Score) while remaining invariant to task-irrelevant environmental variations (Domain Invariance Score). Evaluating 23 encoders, we reveal patterns across architectures, pre-training datasets, and parameter scales. Our findings show that manipulation-pretrained encoders consistently achieve higher Action Scores, CNN-based encoders demonstrate stronger domain invariance than ViTs, and the best-performing models combine both properties, underscoring DIS and AS as complementary predictors of Sim2Real transferability.

Lars Ankile, Zhenyu Jiang, Rocky Duan et al.

Recent advances in behavior cloning (BC) have enabled impressive visuomotor control policies. However, these approaches are limited by the quality of human demonstrations, the manual effort required for data collection, and the diminishing returns from offline data. In comparison, reinforcement learning (RL) trains an agent through autonomous interaction with the environment and has shown remarkable success in various domains. Still, training RL policies directly on real-world robots remains challenging due to sample inefficiency, safety concerns, and the difficulty of learning from sparse rewards for long-horizon tasks, especially for high-degree-of-freedom (DoF) systems. We present a recipe that combines the benefits of BC and RL through a residual learning framework. Our approach leverages BC policies as black-box bases and learns lightweight per-step residual corrections via sample-efficient off-policy RL. We demonstrate that our method requires only sparse binary reward signals and can effectively improve manipulation policies on high-degree-of-freedom (DoF) systems in both simulation and the real world. In particular, we demonstrate, to the best of our knowledge, the first successful real-world RL training on a humanoid robot with dexterous hands. Our results demonstrate state-of-the-art performance in various vision-based tasks, pointing towards a practical pathway for deploying RL in the real world.

Lars Lien Ankile, Kjartan Krange

This paper presents an ensemble forecasting method that shows strong results on the M4 Competition dataset by decreasing feature and model selection assumptions, termed DONUT (DO Not UTilize human beliefs). Our assumption reductions, primarily consisting of auto-generated features and a more diverse model pool for the ensemble, significantly outperform the statistical, feature-based ensemble method FFORMA by Montero-Manso et al. (2020). We also investigate feature extraction with a Long Short-term Memory Network (LSTM) Autoencoder and find that such features contain crucial information not captured by standard statistical feature approaches. The ensemble weighting model uses LSTM and statistical features to combine the models accurately. The analysis of feature importance and interaction shows a slight superiority for LSTM features over the statistical ones alone. Clustering analysis shows that essential LSTM features differ from most statistical features and each other. We also find that increasing the solution space of the weighting model by augmenting the ensemble with new models is something the weighting model learns to use, thus explaining part of the accuracy gains. Moreover, we present a formal ex-post-facto analysis of an optimal combination and selection for ensembles, quantifying differences through linear optimization on the M4 dataset. Our findings indicate that classical statistical time series features, such as trend and seasonality, alone do not capture all relevant information for forecasting a time series. On the contrary, our novel LSTM features contain significantly more predictive power than the statistical ones alone, but combining the two feature sets proved the best in practice.

Lars Lien Ankile, Morgan Feet Heggland, Kjartan Krange

Within the world of machine learning there exists a wide range of different methods with respective advantages and applications. This paper seeks to present and discuss one such method, namely Convolutional Neural Networks (CNNs). CNNs are deep neural networks that use a special linear operation called convolution. This operation represents a key and distinctive element of CNNs, and will therefore be the focus of this method paper. The discussion starts with the theoretical foundations that underlie convolutions and CNNs. Then, the discussion proceeds to discuss some improvements and augmentations that can be made to adapt the method to estimate a wider set of function classes. The paper mainly investigates two ways of improving the method: by using locally connected layers, which can make the network less invariant to translation, and tiled convolution, which allows for the learning of more complex invariances than standard convolution. Furthermore, the use of the Fast Fourier Transform can improve the computational efficiency of convolution. Subsequently, this paper discusses two applications of convolution that have proven to be very effective in practice. First, the YOLO architecture is a state of the art neural network for image object classification, which accurately predicts bounding boxes around objects in images. Second, tumor detection in mammography may be performed using CNNs, accomplishing 7.2% higher specificity than actual doctors with only .3% less sensitivity. Finally, the invention of technology that outperforms humans in different fields also raises certain ethical and regulatory questions that are briefly discussed.

Lars Lien Ankile, Morgan Feet Heggland, Kjartan Krange

The purpose of this paper is to design a solution to the problem of facial recognition by use of convolutional neural networks, with the intention of applying the solution in a camera-based home-entry access control system. More specifically, the paper focuses on solving the supervised classification problem of taking images of people as input and classifying the person in the image as one of the authors or not. Two approaches are proposed: (1) building and training a neural network called WoodNet from scratch and (2) leveraging transfer learning by utilizing a network pre-trained on the ImageNet database and adapting it to this project's data and classes. In order to train the models to recognize the authors, a dataset containing more than 150 000 images has been created, balanced over the authors and others. Image extraction from videos and image augmentation techniques were instrumental for dataset creation. The results are two models classifying the individuals in the dataset with high accuracy, achieving over 99% accuracy on held-out test data. The pre-trained model fitted significantly faster than WoodNet, and seems to generalize better. However, these results come with a few caveats. Because of the way the dataset was compiled, as well as the high accuracy, one has reason to believe the models over-fitted to the data to some degree. An added consequence of the data compilation method is that the test dataset may not be sufficiently different from the training data, limiting its ability to validate generalization of the models. However, utilizing the models in a web-cam based system, classifying faces in real-time, shows promising results and indicates that the models generalized fairly well for at least some of the classes (see the accompanying video).