456 papers - avg viability 6.3
Current research in robotics is increasingly focused on enhancing manipulation capabilities and operational efficiency through advanced learning frameworks. Recent work on force-adaptive reinforcement learning is enabling humanoid robots to maintain balance and adapt to external forces during bimanual tasks, significantly improving their manipulation envelope. Concurrently, generalizable manipulation techniques are being developed to handle diverse object interactions, leveraging task-conditioned frameworks that separate grasping from execution to enhance performance across various object types. Innovations in vision-language-action models are also notable, with frameworks that integrate fine-grained instruction alignment and coarse-to-fine action generation, allowing robots to better understand and execute complex tasks. Additionally, the introduction of adaptive control mechanisms using large language models is bridging the gap between high-level reasoning and low-level execution, improving performance in contact-rich environments. These advancements collectively address pressing commercial challenges in automation, such as optimizing warehouse operations and enhancing the dexterity of robotic systems in dynamic settings.
FAME is a force-adaptive reinforcement learning framework that enhances humanoid manipulation by adapting to external forces in real-time.
An open framework for fine-grained instruction alignment in vision-language-action models, enabling steerable robot policies and improved robotic video understanding.
Scalable autonomous humanoid skills using trainable guided diffusion models trained on diverse motion data.
RLDX-1 is a general-purpose robotic policy for dexterous manipulation that unifies vision-language-action capabilities with advanced system-level design for superior performance on complex real-world tasks.
ExpertGen automates expert policy learning in simulation for scalable sim-to-real transfer in robotics.
A two-stage robotic manipulation framework that decouples grasp from interaction using specialized foundation models for improved generalization across heterogeneous objects.
This work introduces Residual Latent Action (RLA) for visual feature-based world models, enabling faster, more accurate predictions and novel robot learning techniques from offline videos.
CF-VLA optimizes action generation in robotic systems by using a coarse-to-fine method, significantly improving efficiency and success rates.
CoRAL is a modular framework that enables zero-shot planning for contact-rich robotic manipulation by decoupling LLM-based high-level reasoning from adaptive low-level control, outperforming baselines by over 50% success rate.
A Deep Reinforcement Learning framework for real-time joint optimization of order allocation and robot scheduling in warehousing systems, achieving significant efficiency gains.