Research Hub
Papers with a full citation graph - what each paper draws on, and who cites it.
51 papers
WebArena: A Realistic Web Environment for Building Autonomous Agents
Shunyu Yao, Dian Yu, Jason Wei, Yonghui Wu, Nan Du, Izhak Shafran, Karthik Narasimhan
Finding: Results show that LLM-based agents (like GPT-4) can successfully complete about 40% of tasks, which is a significant achievement, but still far from human performance (~90%).
ReAct: Synergizing Reasoning and Acting in Language Models
Jeffrey Zhao, Dian Yu, Shunyu Yao
Finding: ReAct outperformed prior methods on interactive decision-making tasks and improved human interpretability and trustworthiness.
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Ed H. Chi, Jason Wei, Quoc V. Le, Fei Xia, Denny Zhou, Xuezhi Wang, Dale Schuurmans, Brian Ichter, Maarten Bosma
Finding: This method significantly improved accuracy on math, logical reasoning, and complex QA tasks, even for smaller models.
A Review of Sensors and Actuators for Robotics
Mohamed S. Ali, Sarah A. Ahmed
Finding: MEMS-based sensors and electric actuators (e.g., brushless motors) are the most popular choices in modern robotics due to their small size, reasonable cost, and high precision.
Automatic Differentiation in Machine Learning: a Survey
Alexey Andreyevich Radul, Atılım Güneş Baydin, Jeffrey Mark Siskind, Barak A. Pearlmutter
Finding: Reverse-mode AD is the foundational algorithm that powers the training of virtually all deep learning models — commonly known in the ML community as backpropagation.
Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields
Zhe Cao, Tomas Simon, Shih-En Wei, Yaser Sheikh
Finding: This bottom-up architecture achieved high accuracy while maintaining real-time processing speeds, crucially remaining fast regardless of how many people or objects were in the image.
An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition
Cong Yao, Xiang Bai, Baoguang Shi
Finding: The proposed method achieved competitive and superior accuracy on standard benchmarks like IIIT-5K and ICDAR, while being end-to-end trainable and requiring no manual segmentation.
Emergence of Locomotion Behaviours in Rich Environments
Jay Lemay, David Silver, Nicolas Heess, Srinivasan Sriram, Darren King, Koray Kavukcuoglu
Finding: The trained policies exhibited emergent behaviors like walking, running, jumping, and even leaping across various environments, demonstrating robust locomotion.
Attention Is All You Need
Aidan N. Gomez, Łukasz Kaiser, Noam Shazeer, Illia Polosukhin, Ashish Vaswani, Niki Parmar, Jakob Uszkoreit, Llion Jones
Finding: The Transformer achieved state-of-the-art translation quality while being significantly faster to train due to parallelization.
You Only Look Once: Unified, Real-Time Object Detection
Ross Girshick, Ali Farhadi, Santosh Divvala, Joseph Redmon
Finding: YOLO achieved state-of-the-art performance at incredibly high speeds (up to 45 frames per second on standard hardware, and 155 fps for smaller versions), proving that a single end-to-end network could achieve both high accuracy and real-time processing.
Adam: A Method for Stochastic Optimization
Jimmy Ba, Diederik P. Kingma
Finding: Adam combines the advantages of AdaGrad (handles sparse gradients) and RMSProp (works well in non-stationary settings), and performs well empirically across problems.
Human-level control through deep reinforcement learning
Diederik Kingma, Ioannis Antonoglou, Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Martin Riedmiller, Alex Graves
Finding: DQN achieved human-level or better performance on 49 Atari games, marking the first major success of Deep RL.
How transferable are features in deep neural networks?
Jason Yosinski, Yoshua Bengio, Jeff Clune, Hod Lipson
Finding: They found that features become increasingly specific to the original task in higher layers, but still transfer well to related tasks. Transferability degrades as the divergence between tasks increases.
Efficient Estimation of Word Representations in Vector Space
Kai Chen, Greg Corrado, Jeffrey Dean, Tomas Mikolov
Finding: The models learned high-quality word vectors from 1.6 billion words in under a day, orders of magnitude faster than earlier neural approaches. Strikingly, the resulting vector space encoded semantic and syntactic relationships as directions: vector arithmetic such as king − man + woman produced a vector closest to queen, showing that regularities in language were captured as consistent geometric offsets.
ImageNet Classification with Deep Convolutional Neural Networks
Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever
Finding: AlexNet won the ImageNet 2012 competition by a large margin (15.3% top-5 error vs 26.2% for the second best), proving that deep CNNs are highly effective on large-scale data.
An efficient orientation filter for inertial and inertial/magnetic sensor arrays
Sebastian O. H. Madgwick
Finding: The Madgwick filter can operate at high sampling rates (>500 Hz) with low error, achieving performance comparable to extended Kalman filters in practice.
ROS: an open-source Robot Operating System
Brian Gerkey, Ken Conley, Morgan Quigley
Finding: ROS quickly became the industry standard for robot development, accelerating robotics research due to its open-source nature and active community.
The Care‑O‑bot 3 – A Service Robot for Home Environments
Christoph Staehle, Birgit Graf
Finding: Care‑O‑bot 3 was shown to effectively perform tasks like object handling, drink serving, and video communication, receiving positive feedback in user trials.
A survey of robot learning from demonstration
Sonia Chernova, Manuela Veloso, Brett Browning, Brett D. Argall
Finding: They show that learning from demonstration can effectively transfer complex skills to robots, and combining it with RL can yield results surpassing the expert.
Matrix Factorization Techniques for Recommender Systems
Chris Volinsky, Robert Bell, Yehuda Koren
Finding: Matrix factorization yielded superior predictive accuracy compared to nearest-neighbor techniques. By modeling implicit feedback, temporal effects, and confidence levels within the factorization framework, the team successfully won the $1 million Netflix Prize.
Safety in human-robot interaction: a survey
Bruno Siciliano, Alberto De Santis
Finding: They show that combining force/torque sensors with adaptive control can significantly reduce risks, and existing standards (e.g., ISO) provide a solid framework.
The Shadow Dexterous Hand: A High-Fidelity Humanoid Hand for Telemanipulation
Richard Green, Lucas Koch
Finding: The hand was shown to perform delicate tasks like grasping an egg or playing piano, demonstrating its capability as a human assistant in hazardous environments.
Zero-moment point – thirty five years of its life
Branislav Borovac, Miomir Vukobratović
Finding: ZMP has become a cornerstone of bipedal stability control, and many successful algorithms are based on it.
A survey of socially interactive robots
Illah Nourbakhsh, Kerstin Dautenhahn, Terrence Fong
Finding: A unified framework for understanding socially interactive robots and identification of open research issues in HRI.
Random Forests
Leo Breiman
Finding: Random Forest was shown to outperform single trees and other ensemble methods, handling high-dimensional and noisy data well.
The 3D Linear Inverted Pendulum Mode: A simple modeling for a biped walking pattern generation
Shuuji Kajita, Kazuya Yokoi, Kenji Kaneko, Hirohisa Hirakawa, Fumio Kanehiro
Finding: The LIPM model was shown to predict walking motion well for biped robots, and controllers based on it provide stable locomotion.
Random Forests
Leo Breiman
Finding: Generalisation error converges as the number of trees grows; the method is robust to noise and also yields variable-importance measures.
RRT-Connect: An Efficient Approach to Single-Query Path Planning
James J. Kuffner, Steven M. LaValle
Finding: RRT-Connect significantly reduces planning time compared to standard RRT, making it highly effective for single-query problems.
Force control in robotic manipulation: a review
J. Norberto Pires
Finding: Impedance control is better suited for soft environments, while hybrid control is preferred for rigid surfaces; the choice depends on the application.
The PageRank Citation Ranking: Bringing Order to the Web
Sergey Brin, Terry Winograd, Lawrence Page, Rajeev Motwani
Finding: The values in this principal eigenvector represented the steady-state probability of a random web surfer landing on any given page. This metric, 'PageRank', provided an incredibly robust and highly relevant ranking of web pages.
Monte Carlo Localization for Mobile Robots
Wolfram Burgard, Dieter Fox, Frank Dellaert
Finding: The particle filter can handle multi‑modal distributions and outperforms the Kalman filter in real‑world environments with sensor noise.
The development of Honda humanoid robot
Kazuo Hirai, Masato Hirose, Kenichiro Nagasaka
Finding: ASIMO achieved stable walking, stair climbing, and even running, demonstrating that careful mechanical design combined with proper control can produce human-like performance.
Long Short-Term Memory
Sepp Hochreiter, Jürgen Schmidhuber
Finding: LSTM outperformed vanilla RNNs on tasks like speech recognition, machine translation, and text generation, successfully modeling long-term dependencies.
A tutorial on visual servo control
Seth Hutchinson, Gregory D. Hager, Peter I. Corke
Finding: Image-based servoing is generally faster and more robust, but may suffer from field-of-view limitations, while position-based servoing requires accurate 3D estimation but facilitates path planning.
Regression Shrinkage and Selection via the Lasso
Robert Tibshirani
Finding: Lasso was shown to produce more interpretable models compared to Ridge regression and works well in high-dimensional settings.
A study of cross-validation and bootstrap for accuracy estimation and model selection
Ron Kohavi
Finding: 10-fold cross-validation generally offers a good bias-variance trade-off and is often the best choice.
Learning to predict by the methods of temporal differences
Richard S. Sutton
Finding: TD methods were shown to converge faster than Monte Carlo methods in prediction problems and became a cornerstone of modern RL.
A Computational Approach to Edge Detection
John F. Canny
Finding: The Canny edge detector was shown to be statistically optimal and outperformed previous methods, becoming the gold standard for edge detection.
Learning representations by back-propagating errors
Ronald J. Williams, David E. Rumelhart, Geoffrey E. Hinton
Finding: Backpropagation effectively trains multi-layer networks for nonlinear tasks like pattern recognition and representation learning.
Learning representations by back-propagating errors
Ronald J. Williams, Geoffrey E. Hinton, David E. Rumelhart
Finding: Hidden units learn to represent important task features on their own (e.g. 'person', 'generation', 'nationality' in a family-tree problem) — deep networks learn abstract representations, not just memorize.
A new polynomial-time algorithm for linear programming
Narendra Karmarkar
Finding: Proved the algorithm runs in polynomial time, answering the open question, and claimed practical speed advantages over Simplex for large problems.
Gauss and the invention of least squares
Stephen M. Stigler
Finding: Gauss introduced the least squares method in 1809 to solve problems in astronomy and geodesy, and it remains one of the most important statistical methods.
Gaussian Elimination is not Optimal
Volker Strassen
Finding: By recursively applying this 7-multiplication technique, Strassen proved that matrix multiplication could be performed in O(N^2.807) time, fundamentally breaking the assumed O(N^3) barrier and spawning a new field of algebraic complexity theory.
Some methods for classification and analysis of multivariate observations
James MacQueen
Finding: K-Means was shown to be a simple and effective clustering method, though it may converge to local optima.
The regression analysis of binary sequences (with discussion)
David R. Cox
Finding: Logistic regression was shown to be an effective tool for modeling the probability of a binary event based on predictor variables.
A kinematic notation for lower-pair mechanisms based on matrices
Richard S. Hartenberg, Jacques Denavit
Finding: The DH method became a worldwide standard for modeling robot arm kinematics and is used in all industrial and research robots.
Equation of State Calculations by Fast Computing Machines
Edward Teller, Nicholas Metropolis, Marshall N. Rosenbluth, Augusta H. Teller, Arianna W. Rosenbluth
Finding: Demonstrated that sampling from a constructed Markov Chain allows accurate simulation of physical systems and computation of thermodynamic properties.
Optimum Settings for Automatic Controllers
Nathaniel B. Nichols, John G. Ziegler
Finding: The Ziegler-Nichols methods quickly became industry standard for PID tuning and are still used as a good starting point.
On the Problem of the Most Efficient Tests of Statistical Hypotheses
Jerzy Neyman, Egon S. Pearson
Finding: Proved that the likelihood ratio test is the most powerful test for comparing two simple hypotheses at a fixed significance level.
The Probable Error of a Mean
William Sealy Gosset
Finding: Proved that for small samples, the distribution of the mean has heavier tails than the normal distribution, and provided tables of critical values for statistical significance.
An Essay towards Solving a Problem in the Doctrine of Chances
Thomas Bayes
Finding: The derivation of the basic formulation of posterior distributions, laying the foundational framework for what is now known as Bayesian inference.