✷ The archive · 11 labs · 3,065 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
OpenAI Sim-to-Real Transfer of Robotic Control with Dynamics Randomization
OpenAI researchers developed a dynamics randomization method that allows robotic control policies trained exclusively in simulation to transfer to real-world hardware without further physical training.
OpenAI Asymmetric Actor Critic for Image-Based Robot Learning
OpenAI introduces an asymmetric actor-critic framework that improves robot learning by training the critic on full simulator states while the actor relies only on RGBD images, enabling efficient simulation-to-real transfer without real-world data.
OpenAI Domain Randomization and Generative Models for Robotic Grasping
OpenAI developed a data generation pipeline using domain randomization and an autoregressive model to achieve an 80% real-world grasping success rate on unseen objects using only simulated training data.
OpenAI Meta-learning for Wrestling
OpenAI demonstrated that meta-learning agents can quickly adapt their tactics to defeat stronger fixed-policy agents and recover from physical malfunctions in simulated robot wrestling.
OpenAI Competitive Self-Play Research
OpenAI demonstrated that competitive self-play allows simulated 3D robots to autonomously discover complex physical skills like tackling and diving without explicit human design.
Nonlinear Computation in Deep Linear Networks
OpenAI researchers discovered that deep linear networks can perform nonlinear computation by exploiting floating-point arithmetic underflow around zero.
OpenAI and Oxford Introduce LOLA for Multi-Agent Reinforcement Learning
OpenAI and the University of Oxford have developed Learning with Opponent-Learning Awareness (LOLA), an RL agent that shapes the learning of other agents to achieve mutually beneficial outcomes.
OpenAI Learning with Opponent-Learning Awareness (LOLA)
OpenAI introduces Learning with Opponent-Learning Awareness (LOLA), a multi-agent reinforcement learning method that enables agents to shape the learning of other agents to foster cooperation and reach Nash equilibrium.
OpenAI Baselines: ACKTR & A2C
OpenAI has released Baselines implementations of ACKTR and A2C, providing a more sample-efficient reinforcement learning algorithm (ACKTR) and a synchronous, deterministic variant of A3C (A2C).
OpenAI Dota 2 1v1 Bot Results
OpenAI developed a reinforcement learning agent that achieved superhuman performance in Dota 2 1v1, demonstrating that self-play can scale machine learning systems beyond human-level capability.
OpenAI Dota 2 Bot Announcement
OpenAI has developed a bot capable of defeating top professional Dota 2 players in 1v1 matches using self-play reinforcement learning without imitation learning or tree search.
OpenAI RL-Teacher Release
OpenAI has released RL-Teacher, an open-source interface that allows AI agents to be trained via occasional human feedback instead of hand-crafted reward functions.
OpenAI Better Exploration with Parameter Noise
OpenAI introduces a method of adding adaptive noise to neural network parameters rather than the action space, significantly boosting reinforcement learning performance and exploration consistency.
Proximal Policy Optimization (PPO) Release Notes – OpenAI Baselines
OpenAI released Proximal Policy Optimization (PPO), a simple-to-implement reinforcement learning algorithm that matches or exceeds state-of-the-art performance while becoming the default RL method at OpenAI.
OpenAI Robust Adversarial Inputs Research
OpenAI has developed images that remain adversarial across varied scales and perspectives, challenging the notion that multi-perspective image capture in systems like self-driving cars prevents malicious trickery.
OpenAI Hindsight Experience Replay
OpenAI introduced Hindsight Experience Replay, a technique that enables sample-efficient reinforcement learning from sparse, binary rewards by treating failed attempts as successful achievements of alternative goals.
OpenAI Teacher–Student Curriculum Learning
OpenAI introduces Teacher–Student Curriculum Learning (TSCL), an automatic framework that optimizes training by selecting subtasks based on the student's learning progress and performance decay.
mujoco-py 1.50.1.0 release notes / what's new
OpenAI has open-sourced mujoco-py 1.50.1.0, a high-performance Python 3 binding for the MuJoCo engine that introduces batched simulation and GPU-accelerated rendering.
OpenAI Learning from Human Preferences
OpenAI and DeepMind developed a reinforcement learning algorithm that infers goals from human preference comparisons, reducing the need for manually written reward functions.
OpenAI MADDPG: Learning to Cooperate, Compete, and Communicate
OpenAI introduced MADDPG, a reinforcement learning algorithm that uses centralized learning and decentralized execution to enable multiple agents to collaborate and compete in complex environments.
OpenAI UCB Exploration via Q-Ensembles
OpenAI researchers developed an exploration strategy using an ensemble of Q*-functions based on Upper-Confidence Bounds (UCB) to improve deep reinforcement learning performance on Atari benchmarks.
OpenAI Baselines: DQN release and reinforcement learning best practices
OpenAI has open-sourced OpenAI Baselines, starting with a high-performance implementation of DQN and three of its variants to improve reproducibility in reinforcement learning.
OpenAI Robots That Learn: One-Shot Imitation Learning from Simulation
OpenAI has developed a robotics system trained entirely in simulation that can learn new tasks from a single human demonstration provided via VR.
OpenAI Roboschool Release
OpenAI has released Roboschool, a collection of physics-based environments for reinforcement learning that ports MuJoCo tasks to the Bullet physics engine and introduces new challenges in interactive control and multiplayer training.
Equivalence between policy gradients and soft Q-learning
OpenAI researchers demonstrate that soft (entropy-regularized) Q-learning is mathematically equivalent to policy gradient methods, providing a theoretical bridge between two primary model-free reinforcement learning approaches.
Stochastic Neural Networks for Hierarchical Reinforcement Learning
OpenAI researchers propose a framework using Stochastic Neural Networks and information-theoretic regularizers to pre-train interpretable skills that accelerate learning in downstream tasks with sparse rewards.
OpenAI Unsupervised Sentiment Neuron
OpenAI developed an unsupervised system that learns sentiment representation by predicting the next character in Amazon reviews, achieving state-of-the-art accuracy on the Stanford Sentiment Treebank with significantly fewer labeled examples.
OpenAI Spam Detection in the Physical World
OpenAI developed a robotic system capable of detecting physical cans of Spam using a VGG16-based neural network trained entirely in simulation via domain randomization.
Evolution Strategies as a Scalable Alternative to Reinforcement Learning
OpenAI researchers demonstrate that Evolution Strategies (ES) can rival the performance of standard reinforcement learning on Atari and MuJoCo benchmarks while offering superior scalability and simpler implementation.
OpenAI One-Shot Imitation Learning
OpenAI introduces a meta-learning framework for one-shot imitation learning that allows robots to learn new tasks from a single demonstration and generalize to new situations without task-specific engineering.
OpenAI Support for Distill Journal Launch
OpenAI announced its support for Distill, a new journal utilizing modern web technologies to improve the communication of machine learning results.
Learning to communicate: OpenAI's 2017 research on emergent AI language
OpenAI's 2017 'Learning to communicate' work shows how reinforcement‑learning agents can invent a grounded, compositional language to solve cooperative tasks in simple simulated worlds.
Emergence of grounded compositional language in multi-agent populations
OpenAI researchers showed that grounded compositional language can emerge in multi-agent populations, demonstrating how agents develop a structured symbol-based communication system to achieve goals.
OpenAI Prediction and Control with Temporal Segment Models
OpenAI introduces a deep generative model that predicts future state trajectories over temporal segments rather than discrete timesteps, enabling stable long-horizon predictions for complex nonlinear systems.
OpenAI Third-Person Imitation Learning
OpenAI introduces an unsupervised third-person imitation learning method that allows agents to learn tasks from demonstrations provided from a different viewpoint without explicit state correspondence.
Attacking Machine Learning with Adversarial Examples
OpenAI explores how adversarial examples—intentionally designed inputs that cause ML models to fail—highlight critical vulnerabilities in supervised and reinforcement learning, posing significant challenges for AI safety.
Adversarial Attacks on Neural Network Policies
OpenAI researchers demonstrate that neural network policies in reinforcement learning are vulnerable to small, imperceptible adversarial perturbations that significantly degrade performance across various tasks and training algorithms.
OpenAI Team Update January 2017
OpenAI expanded its team to 45 members in January 2017, adding specialists in robotics, distributed systems, and optimization to advance AI capabilities.
PixelCNN++: Improving PixelCNN with discretized logistic mixture likelihood and other modifications
OpenAI released PixelCNN++, an enhanced PixelCNN implementation that replaces the 256-way softmax with a discretized logistic mixture likelihood, conditions on whole pixels, adds downsampling, shortcut connections, and dropout, achieving state-of-the-art log likelihood on CIFAR-10.
OpenAI: Faulty Reward Functions in the Wild
OpenAI identifies the risk of reward misspecification in reinforcement learning, where agents exploit imperfect proxies to achieve high scores through unintended behaviors.
OpenAI Universe
OpenAI Universe is a general-purpose AI training platform that allows agents to interact with any computer program via screen pixels and virtual keyboard/mouse inputs to foster the development of general intelligence.
OpenAI and Microsoft Partnership Announcement
OpenAI has partnered with Microsoft to use Azure as its primary cloud platform for deep learning and AI research to accelerate the scale of its experiments.
OpenAI Study on Count-Based Exploration for Deep Reinforcement Learning
OpenAI researchers have demonstrated that a simple generalization of classic count-based exploration, using hash codes to map high-dimensional states to counts, can achieve near state-of-the-art performance in deep reinforcement learning.
On the quantitative analysis of decoder-based generative models
OpenAI researchers propose using Annealed Importance Sampling to accurately evaluate log-likelihoods in decoder-based generative models, addressing the difficulty of quantifying performance beyond simple sample inspection.
OpenAI: A Connection Between GANs, Inverse Reinforcement Learning, and Energy-Based Models
OpenAI researchers demonstrate that certain Inverse Reinforcement Learning (IRL) methods, specifically maximum entropy IRL, are mathematically equivalent to Generative Adversarial Networks (GANs) and Energy-Based Models (EBMs).
OpenAI RL²: Fast Reinforcement Learning via Slow Reinforcement Learning
OpenAI introduces RL², a method that uses a recurrent neural network (RNN) to learn a reinforcement learning algorithm, allowing agents to adapt to new tasks in a few trials by encoding the learning process itself into the RNN weights.
OpenAI Variational Lossy Autoencoder Research
OpenAI introduces a Variational Lossy Autoencoder that combines VAEs with neural autoregressive models to learn global representations while discarding irrelevant local details like image texture.
OpenAI Neural GPU: Extensions and Limitations
OpenAI researchers have improved the Neural GPU's ability to learn algorithmic tasks like decimal arithmetic and complex expressions by using curriculum learning and increased model size.
OpenAI Semi-supervised Knowledge Transfer for Private Training Data
OpenAI introduces Private Aggregation of Teacher Ensembles (PATE), a framework that uses noisy voting among teacher models to train a student model with strong differential privacy guarantees.
OpenAI Self-Organizing Conference on Machine Learning (SOCML)
OpenAI hosted its first self-organizing conference on machine learning to accelerate AI research through peer-to-peer education and serendipitous interaction among 150 AI practitioners.