Async RL fine-tuning via EXPO-FT enables π0.5 to handle highly dynamic manipulation tasks such as ball balancing and striking in real time. The approach demonstrates that post-training with RL can unlock dynamic capabilities not achievable through imitation learning alone.

Aether's unified world model jointly learns environment dynamics and long-term physical state evolution under physical constraints, enabling robots to plan feasible trajectories over extended horizons. The demo highlights how constraint-aware world modeling can prevent physically implausible plans that plague standard generative approaches.

Proposes a framework that restores the Markov property for RL fine-tuning of VLAs despite asynchronous inference delays, enabling effective RL training on large vision-language-action models. This directly unblocks post-training RL for frontier robot policies operating under real-world latency constraints.

A 6B-parameter world-action model that jointly predicts 3D pointmaps and DINO features alongside RGB frames, improving demonstration efficiency, generalization, and long-horizon task performance in robotics. Integrating structured 3D world predictions into the action loop represents a meaningful advance in robot foundation model design.

SkipVLA proposes a hybrid architecture combining a classical planner with a Vision-Language-Action model for manipulation tasks, achieving 2.5x faster execution and 59.6% lower latency on a physical robot arm. The approach demonstrates that structured planning can significantly reduce the computational burden of end-to-end VLA inference.

ForeDrive introduces a planning-relevant latent world model that guides autonomous driving through foresight, enabling the agent to anticipate future states before acting. The approach advances world-model-based driving by tightly coupling latent prediction with planning objectives.

A new algorithm reconstructs and retargets monocular RGB (or synthetically generated) videos to high-DoF robot hands, enabling scalable dexterous manipulation data collection that outperforms current state-of-the-art retargeting methods. This matters because it removes the need for expensive motion capture or teleoperation hardware to generate training data for dexterous robots.

GAM is a robot foundation model that conditions action chunk prediction on 3D grounded inputs — including language, 3D point clouds, and bounding boxes — alongside robot history, achieving strong results on LIBERO benchmarks and real-robot tasks. The 3D grounding mechanism advances how embodied agents spatially understand and act in their environment.

Autonomous navigation in complex, unstructured environments poses a significant challenge, with traditional planners lacking adaptability and end-to-end learning methods hindered by data dependency or training instability. This paper proposes a hierarchical learning-based hybrid method with sim-to-real transfer.

This work distills frozen world-model representations into compact VLAs via a single feature-alignment loss at training time, with the teacher cached once and absent at inference; a 0.8B student achieves 97.9% on LIBERO and lifts RoboCasa-GR1 humanoid performance from 48.2% to 50.5% at zero extra runtime cost (32 ms / 1.86 GB on RTX 5090). The approach generalizes across scales and architectures on both single-arm and bimanual real hardware.