Danijar Hafner is developing AI agents that can plan for unexpected situations by navigating environments they have not previously encountered in training. He utilizes model-based reinforcement learning to create world models that emulate physical reality for these agents. This method allows agents and robots to operate in unfamiliar spaces by making predictions about future outcomes. Hafner's work has involved training agents to solve complex tasks in video games before migrating them to physical robotics. His goal is to enable robots to navigate human spaces by handling novel floor plans and furniture.
A video-augmented virtual reality framework is presented to enhance passenger understanding of physical autonomous vehicle behavior. This system couples a physical ROS 2 robot vehicle with a Unity 6 application on a Meta Quest 3S headset. The framework streams the vehicle's state and live camera feed to allow the VR to mirror the physical robot's motion and display navigation decisions. Evaluation over 20 trials showed low latency and accurate representation of route progress between the physical and virtual vehicles. The results confirm the framework's ability to provide temporally synchronized and semantically consistent observation of autonomous vehicle actions.
Vision-Language-Action (VLA) models lack robust safety guarantees when fine-tuned, as current methods often result in constraint violations. ShieldVLA introduces a safety-aligned fine-tuning framework utilizing Hamilton-Jacobi reachability to learn the safe operating region from visual observations. This framework employs a safety critic to gate policy optimization, separating reward maximization from recovery near unsafe states. Rubric-based safety scores are used to provide scalable supervision by converting semantic safety feedback into critic targets. Experiments show ShieldVLA reduces cumulative safety cost by 57% and improves task success rates over existing methods.
A human-in-the-loop interactive reinforcement learning framework is proposed to enable sample-efficient, high-precision robotic assembly in industrialized construction. The framework learns expertise from offline teleoperated demonstrations, sparse event-driven takeovers at contact-failure points, and acceptance-aligned rewards. It utilizes a temporally abstract policy based on Q-chunking with Flow Q-Learning to manage multimodal recovery maneuvers. Evaluation in MuJoCo demonstrated that the pipeline achieved 100% autonomous seating with minimal human supervision over time. The method shows significant sample efficiency, requiring only 12-15 minutes of cumulative supervision for 3.0 hours of online training.
The work learns manipulation-sufficient representations by employing an action-conditioned outcome bottleneck. This bottleneck uses a marginal outcome log-loss and a KL term to regularize a stochastic representation, motivated by preserving the outcome distribution of admissible actions. The resulting model is evaluated as a rate-regularized mixture predictor. The method supports various manipulation tasks and exhibits strong performance in grasping, achieving an AUC of 0.876 on simulated data. This approach leverages a finite-probe theorem to relate local tangent spaces to the outcome Jacobian.
Self-evolving agents aim to improve from post-deployment experience, addressing the issue where deployed humanoid systems remain static. The paper examines self-evolution mechanisms for embodied agents, defining a state tuple including policy, memory, and body. Four mechanisms of self-evolution are presented: self-learning, self-adaptation, self-optimization, and self-generation. Safety and uncertainty are incorporated into the evolution operator via a world-model verification gate within a human-oversight envelope. Evaluation should track the evolving trajectory rather than fixed checkpoints, highlighting the need for new benchmarks.
GzDRL is a new single-process reinforcement learning framework for Gazebo designed to improve reproducibility and scalability in robotics experimentation. It introduces a middleware-free environment-stepping mechanism to directly synchronize agent actions and physics updates deterministically. This design allows for high-throughput data collection and reproducible RL training and evaluation. Benchmarks show GzDRL achieves high workstation throughput while maintaining synchronization and scalability. The method also successfully validates sim-to-real transfer by deploying learned policies to physical hardware.
The research investigates dark manipulation, aiming to achieve contact-rich skills using only visual latents derived after initial lighting. The proposed method, Physical Kernel, uses structured visual latents for this manipulation. Experiments show that the dark manipulation method achieves significant stacked success on stacking tasks compared to traditional methods. The study analyzes how occlusion, action shuffling, and writing time affect the encoding process for these skills. Furthermore, the work explores the performance of different observation methods in achieving dark manipulation goals.
A conflict-predictive variable horizon is proposed for distributed model predictive control in multi-drone collision avoidance. This method allows each drone to locally set a prediction horizon based on extrapolated neighbor flight lines and conflict tests. The horizon dynamically adjusts to be the smallest value covering the farthest predicted conflict, collapsing when airspace is clear. The method preserves recursive feasibility and asymptotic stability under computable feasibility bounds for linear and practical nonlinear models. Simulation results show this approach reduces computational cost while maintaining drone separation compared to fixed horizons.
The research proposes a framework to extend the speed limit of quadrupedal locomotion by addressing limitations in existing methods. It introduces a refined actuator model to accurately capture high-speed voltage coupling and magnetic saturation, improving sim-to-real performance. A reinforcement learning approach with adaptive command scheduling stabilizes training over a wide command distribution. Experiments on a BlackPanther2 robot achieved speeds of up to 13.2 m/s on a treadmill and 11.65 m/s outdoors. The findings emphasize the necessity of accurate actuator modeling for safe policy learning and show the benefit of adaptive scheduling.
Superior quadrupedal locomotion can be achieved by exploiting intrinsic dynamics, rather than solely increasing actuator capacity. Inter-limb inertial coupling redistributes energy across the gait cycle, which reduces peak joint torque needed for rapid motion. Incorporating hardware parameters into the analysis allows for the systematic utilization of this inertial coupling in robot design. This approach enabled a quadruped robot to achieve a running speed of 10.74 m/s, surpassing 10 m/s. These findings highlight inertial coupling as a key mechanism for high-speed legged locomotion.