Embodied control work concentrates on grounded decision mechanisms

Memory is the main concrete mechanism in the day’s embodied control paper. The UAV system turns a scene into semantic events, retrieves similar past situations from a knowledge bank, then picks an action from the best-matching maneuver cluster. The paper’s claim is practical: keep the control loop fast enough for deployment while leaving a trace of why an action was chosen. Reported results are strong inside its own setup, with 20–50 ms control intervals, sub-millisecond retrieval, and 100% success with zero collisions across five adversarial curriculum episodes. The evidence is narrower than a benchmark-heavy robotics paper, since the excerpt does not show external baselines or broad ablations.

The reinforcement learning paper centers on keeping imagined trajectories tied to reality over long horizons. GIRL adds a grounding signal from DINOv2, a vision model, and replaces a fixed KL term with an adaptive trust-region bottleneck that reacts to uncertainty and real-transition feedback. The reported gains are broad: IQM 0.81 on DeepMind Control versus 0.67 for DreamerV3 and 0.71 for TD-MPC2, plus lower drift scores and better results on distractor and manipulation tasks. A distilled variant also cuts the grounding overhead from 22% of wall-clock time to under 4%, which matters if the method is to stay competitive in training speed.