This paper introduces FoAM, a novel MTIL policy that conditions on multi-modal goals (e.g. language instruction + goal image) and — crucially — predicts the visual consequences (future states) of its actions (foresight). By training both to reconstruct expert actions and to align predicted future states with the goal, FoAM equips robotic manipulation agents with more expressive, forward-looking embeddings. Evaluated over 100+ manipulation tasks in simulation and real world, FoAM achieves up to a 41% improvement in success rate over prior state-of-the-art. The authors also release a simulation benchmark: a dual-arm system replicating a UR3e robot, with 10 scenario suites totaling over 80 tasks for training and evaluation.

This paper proposes a novel 'Lazy A* with Motion Primitives' algorithm tailored for UAVs. By discretizing control inputs into motion primitives (short-duration dynamically feasible motions) and delaying expensive collision and feasibility checks until necessary, the method significantly reduces computation time. The authors also improve the sampling of control inputs via a normal distribution to enhance state-space coverage. Experiments show that the approach finds optimal or near-optimal trajectories while being efficient enough for onboard real-time planning.