Aims to create a world model without baggage, achieving good performance.
problem Creating world models without baggage and limitations.
method Self-supervised representation learning, frame and action stacking, data augmentation.
result Good performance on the Atari 100k benchmark.
DreamerV2 learns Atari game behaviors from a world model, achieving human-level performance.
problem Learning complex behaviors in Atari games from limited data.
method DreamerV2 uses a world model with discrete representations to predict behaviors in a compact latent space.
result Achieves human-level performance on 55 Atari tasks.
Agents need world models to generalize multi-step tasks.
problem The necessity of world models for flexible, goal-directed behavior.
method Formal analysis and demonstration of the necessity of world models for agents to generalize multi-step tasks.
result World models are necessary for agents to generalize to multi-step goal-directed tasks.
Dreamer 4 learns Minecraft tasks from videos alone.
problem Accurately predicting object interactions in complex environments.
method Reinforcement learning inside a fast, accurate world model.
result Dreamer 4 outperforms previous models in Minecraft, learning from only offline data.
FOCUS improves offline RL by incorporating causal structure into world-models.
problem Learning effective policies from historical data without interaction.
method FOCUS proposes a practical algorithm that learns and leverages causal structure in offline RL.
result FOCUS outperforms plain model-based offline RL algorithms and other causal model-based RL algorithms.
Unified model improves object-centric world modeling with new abilities.
problem Integration of recent advances and temporal imagination abilities.
method Generative Structured World Models (G-SWM) framework.
result G-SWM achieves best or comparable performance in temporal generation.
This work designs an active world model learning system with progress-based curiosity.
problem Learning compact world models from visual exploration.
method Constructs a curious agent in a 3D environment, using γ-Progress as a curiosity signal. result The γ-Progress-driven controller outperforms state-of-the-art exploration strategies. We explore building generative neural network models of popular reinforcement learning environments. Our world model can be trained quickly in an unsupervised manner to learn a compressed spatial and temporal representation of the environment. By using features extracted from the world model as inputs to an agent, we c…
WMPG reduces policy gradient variance using world models.
problem Reducing variance in policy gradient estimates.
method Trains a world model online to estimate policy gradients and uses imagined trajectories as a baseline.
result WMPG achieves better sample efficiency compared to AC and MAC.
A new RL algorithm POWR learns world models to estimate action-values.
problem Inaccessibility of explicit action-value functions in RL.
method Learning a world model using conditional mean embeddings and deriving action-value function via matrix operations.
result POWR algorithm converges to global optimum with proven rates.
Curiosity-Critic improves world model training by focusing on cumulative prediction error.
problem Training world models with intrinsic rewards that consider cumulative prediction error.
method Curiosity-Critic uses a surrogate reward based on the difference between current and asymptotic prediction errors, estimated online by a co-trained critic.
result Curiosity-Critic outperforms other methods in training speed and final world model accuracy.
SPARTAN learns sparse interaction graphs between objects in scenes.
problem Capturing structured interactions between entities in complex settings remains challenging.
method SPARTAN is a Transformer-based world model that learns sparse, context-dependent interaction graphs through sparsity regularization on attention patterns.
result SPARTAN achieves significantly improved few-shot adaptation to dynamics changes and robustness against distractors.
New method accelerates MBRL by 10x for complex quadrotor tasks.
problem Inefficient training of world models in MBRL for complex robotic tasks.
method Parallelized state-space models and privileged information for training.
result Significant speedup in world model and MBRL training times.
Transformer learns long-term dependencies from real-world data.
problem Sample inefficiency in deep reinforcement learning.
method Transformer architecture applied to autoregressive real-world episodes.
result Transformer-based world model generates meaningful experience.
A generative recurrent neural network is quickly trained in an unsupervised manner to model popular reinforcement learning environments through compressed spatio-temporal representations. The world model's extracted features are fed into compact and simple policies trained by evolution, achieving state of the art resul…
Paper presents a world model that learns invariant causal features using contrastive unsupervised learning.
problem Learning invariant causal features in unsupervised settings.
method Contrastive unsupervised learning with intervention invariant auxiliary task.
result Significantly outperforms state-of-the-art methods on out-of-distribution point navigation tasks.
A new neural network model improves reinforcement learning efficiency.
problem Improving sample efficiency in reinforcement learning.
method Proposes a neural network architecture using VQ-VAE and convolutional LSTM for world modeling.
result Shows comparable performance to SimPLe with significantly smaller model.
Unified definition of hallucinations in language models.
problem Persistent hallucinations despite mitigation efforts.
method Unified definition of hallucination as inaccurate world modeling.
result Unified framework distinguishes hallucinations from other errors.
Paper tackles continual reinforcement learning by forgetting, proposing a planning method with online world models.
problem Catastrophic forgetting in reinforcement learning when learning new tasks.
method Planning with an online world model using model predictive control.
result The proposed FTL Online Agent (OA) learns new tasks without forgetting old skills.
New model learns causal world dynamics from state space models.
problem Lack of causal world models in neural world modeling.
method State Space Models (SSM) with attention mechanisms.
result SSM can learn causal models of environments with equivalent performance.
ReCoRe learns invariant features for world navigation using contrastive learning and regularizers.
problem Limited sample efficiency and overfitting to training scenarios in RL for visual navigation.
method Contrastive unsupervised learning and intervention-invariant regularizer.
result Significantly improves sample efficiency and generalization in out-of-distribution point navigation tasks.
UWM-JEPA predicts future scenarios in belief space, improving accuracy in partially observed environments.
problem Predicting future scenarios in partially observed environments with uncertainty.
method Introduces UWM-JEPA, a JEPA world model with a density-matrix latent and learned unitary predictor.
result UWM-JEPA achieves 0.77 accuracy on a hidden-velocity indicator task, outperforming LSTM-JEPA.
This work improves AI's ability to solve physical tasks by optimizing world models in abstracted spaces.
problem Developing AI agents capable of solving diverse physical tasks and generalizing to new environments.
method Investigates and optimizes a family of joint-embedding predictive world models (JEPA-WMs) for efficient planning in abstracted spaces.
result Proposes a model that outperforms two established baselines in both navigation and manipulation tasks.
Infants are experts at playing, with an amazing ability to generate novel structured behaviors in unstructured environments that lack clear extrinsic reward signals. We seek to mathematically formalize these abilities using a neural network that implements curiosity-driven intrinsic motivation. Using a simple but ecolo…
The utility of learning a dynamics/world model of the environment in reinforcement learning has been shown in a many ways. When using neural networks, however, these models suffer catastrophic forgetting when learned in a lifelong or continual fashion. Current solutions to the continual learning problem require experie…
The study proves necessary conditions for robust decision-making in uncertain environments.
problem Conditions for robust decision-making in uncertain environments.
method Quantitative selection theorems and binary betting decisions.
result World models, belief-like memory, and persistent variables are necessary for strong task performance.
Cosmos models scenes using neural encodings and symbolic attributes for compositional generalization.
problem Modeling scenes with high performance on unseen input scenes composed of known visual elements.
method Neurosymbolic grounding with neurosymbolic scene encodings and attention mechanisms.
result Establishes a new state-of-the-art for compositional generalization in world modeling.
Future autonomous systems need reliable world models and complex action sequences.
problem Current automated systems lack reliable world models and complex action sequences.
method Introduce energy-based and latent variable models combined in a hierarchical joint embedding predictive architecture (H-JEPA).
result Combining energy-based and latent variable models in H-JEPA can lead to reliable world models and complex action sequences.
New theorems show agents need specific internal structures to perform well under uncertainty.
problem How do agents need to be structured to perform well under uncertainty?
method Proved selection theorems showing strong task performance forces specific internal structures.
result Strong task performance forces world models, belief-like memory, and persistent regime-tracking variables.
GPT learns a causal world model from token predictions, validated in game sequences.
problem Does GPT implicitly learn a causal world model from token predictions?
method Derived a causal interpretation of GPT's attention mechanism and proposed zero-shot causal structure learning.
result GPT can generate legal next moves with high confidence for sequences with encoded causal structures, but fails for illegal moves.
Agents compose pre-trained policies for complex tasks, improving zero-shot performance.
problem Challenges in long-horizon predictions and estimating visitation distributions induced by policy sequences.
method Learn predictive jumpy world models of multi-step dynamics, enhancing predictions with a consistency objective.
result Compositional planning with jumpy world models yields, on average, a 200% relative improvement over primitive actions on long-horizon tasks.
Equivariant MuZero improves generalization in procedurally generated environments.
problem Leveraging world models for robust generalization in real-world scenarios.
method Incorporates symmetries of the environment into MuZero's world-model architecture, making MuZero equivariant to certain groups of transformations.
result Improves data efficiency and generalization capabilities, demonstrated on procedurally generated MiniPacman and Chaser.
A structured understanding of our world in terms of objects, relations, and hierarchies is an important component of human cognition. Learning such a structured world model from raw sensory data remains a challenge. As a step towards this goal, we introduce Contrastively-trained Structured World Models (C-SWMs). C-SWMs…
A method for self-supervised representation learning in partially observable environments.
problem Sparse rewards and stochasticity in partially observable environments.
method World model in latent space for estimating missing information.
result Significant improvement in exploration compared to prior work.
Infants are experts at playing, with an amazing ability to generate novel structured behaviors in unstructured environments that lack clear extrinsic reward signals. We seek to replicate some of these abilities with a neural network that implements curiosity-driven intrinsic motivation. Using a simple but ecologically …
Integrates skills and world models for efficient task solving and transfer.
problem Quickly solve new tasks in complex environments using reusable knowledge.
method Leverages partial amortization for fast adaptation and online skill planning.
result Improved sample efficiency in single tasks and transfer between tasks.
Proposes DGCN with trajectory sampling for data-efficient policy search in MBRL.
problem Improving data efficiency in model-based reinforcement learning.
method Combines trajectory sampling and DGCN for uncertainty propagation in probabilistic world models.
result Improves sample-efficiency over other uncertainty propagation methods and probabilistic models.
Symbolic grounding in causal dynamics achieves near-infinite temporal consistency.
problem Achieving linear identifiability in non-Gaussian physical systems.
method Physics-Grounded Symbolic Architecture (PGSA)
result PGSA achieves exact linear identifiability for all physical regimes.
Study on RL on volatility surfaces, proving no free lunch for law-seeking methods.
problem Aligning RL agents with no-arbitrage laws in volatile markets.
method Built a law manifold, defined penalties, and used a Goodhart decomposition.
result No free lunch theorem: Law-seeking RL cannot outperform baselines.
LeJEPA learns latent variables from nonlinear observations.
problem Learning latent variables from nonlinear observations.
method Proves linear identifiability of Gaussian latent distributions.
result Gaussian distribution uniquely guarantees linear identifiability.
LEXA learns to discover and achieve goals in unseen environments.
problem Learning to solve diverse tasks in complex visual environments without supervision.
method LEXA learns a world model from image inputs and uses it to train an explorer and an achiever policy from imagined rollouts.
result LEXA solves tasks specified as goal images zero-shot without additional learning.
New method dynamically adjusts UTD ratio to balance under- and overfitting in RL.
problem Balancing under- and overfitting in world model learning for RL.
method Dynamic adjustment of UTD ratio based on validation performance on a small subset of experience data.
result Our method improves balance between under- and overfitting compared to default settings and competitive with extensive hyperparameter search.
We analyze the information-theoretic limits for the recovery of node labels in several network models. This includes the Stochastic Block Model, the Exponential Random Graph Model, the Latent Space Model, the Directed Preferential Attachment Model, and the Directed Small-world Model. For the Stochastic Block Model, the…
Sepsis is a life-threatening condition caused by the body's response to an infection. In order to treat patients with sepsis, physicians must control varying dosages of various antibiotics, fluids, and vasopressors based on a large number of variables in an emergency setting. In this project we employ a "world model" m…
Scalability in terms of object density in a scene is a primary challenge in unsupervised sequential object-oriented representation learning. Most of the previous models have been shown to work only on scenes with a few objects. In this paper, we propose SCALOR, a probabilistic generative world model for learning SCALab…
Deep Neural Networks (DNNs) are usually over-parameterized, causing excessive memory and interconnection cost on the hardware platform. Existing pruning approaches remove secondary parameters at the end of training to reduce the model size; but without exploiting the intrinsic network property, they still require the f…
Soft geometric bias improves physical dynamics predictions.
problem Learning physical dynamics with exact group equivariance can degrade performance.
method Object-centric world models using geometric algebra neural networks.
result Soft geometric inductive bias leads to better physical fidelity predictions.
Agents solving multi-task classification learn disentangled representations.
problem Creating interpretable world models from multi-task data.
method Optimal multi-task classification tasks, noise, evidence accumulation.
result Disentangled representations emerge in multi-task learning.