UCLA

Robots in human environments face a physical gap between their embodiments and human-oriented spaces, and a knowledge gap between task descriptions and the information needed for execution. This dissertation examines these gaps through four complementary studies, with emphasis on human--robot co-activity rearrangement and Semantically Unified Programs (SUN).The physical gap is approached by adapting robot motion to an existing environment or rearranging the environment itself. The Virtual Kinematic Chain (VKC) study includes the author's sampling-based extension for coupled robot--object motion planning. Co-activity rearrangement treats human spatial organization and robot accessibility as joint design objectives. It combines semantic knowledge and learned spatial relations with robot-specific traversability, reachability, and interaction regions to optimize furniture layouts while retaining functional groups. Layouts account for robot footprint, reach, and intended activities. Across 154 SUNCG rooms sampled from a held-out partition, the method reports a 14 percent improvement in its robot-access measure and 30 percent more reachable objects.The knowledge gap is examined at two scales. In tool use, the author's initial trajectory-optimization formulation maps the desired physical effect of a single interaction to a robot-specific strategy. SUN addresses long-horizon, multistage manipulation by grounding language into executable, stage-wise geometric relations shared by transition predicates, model-predictive-control costs, and reinforcement-learning rewards. Model predictive control verifies programs before policy learning. Across five runs per task, repair increased program acceptance from 64.4 to 95.6 percent. The primary stage-conditioned policy with bounded residual learning reached 82.03 percent task-macro success, compared with 83.76 percent for the verification controller; three independent lineages averaged 79.43 percent.The SUN controller produced 246.79 successful trajectory-minutes per GPU-hour,10.57 times the human rate, and broke even near 513 demonstrations. With 500 trajectories per task, its DP3 policy reached 46.02 percent task-macro success. Zero-shot DP3 policies produced 36 successes in 106 physical trials (33.96 percent pooled success) and 34.72 percent task-macro success, with nonzero success on all tested tasks.Together, these studies address physical compatibility and explicit task-knowledge representation as complementary requirements for robot operation: enabling action in shared environments and specifying the physical outcomes that action must achieve.