AI Agent Papers2026
Combines long-horizon task planning with native keyboard-and-mouse control to complete multi-stage Minecraft tasks from visual feedback alone. Structured memory and policy-aware completion checks help the agent verify progress, recover, or replan.
Connects scene understanding, long-horizon visual planning, and robot control by alternating language reasoning with video prediction. An agentic variant learns when generated visual evidence is worth using.
Simulates closed-loop interactions between people and wearable AI assistants to test whether multimodal models can guide real-world tasks. Its 120 multi-turn benchmark tasks are scored against verifiable simulator states.
Rewinds world-model trajectories to test alternative robot actions and turn outcome comparisons into action-level training signals. Real-robot data continually corrects the model; across three physical tasks, success rates beat the strongest baseline by 10.67–11.85 percentage points.
Conditions robot action generation on predicted visual and motion features, capturing physical dynamics without generating future video during control. Reports results across four simulation benchmarks.