Boxi Xia

We present MARBLE, a fully enclosed omnidirectional amphibious rolling robot driven entirely by internal mass redistribution. Three orthogonal linear sliders shift internal masses to roll a sealed spherical shell that serves as ground contact, buoyant hull, and mount for passive fins, so the same mechanism rolls on land and propels itself on water without reconfiguration, and tolerates direct contact with terrain and obstacles.

We introduce the Cartesian Hand, a 7-DoF end-effector that combines independent grasping and relative manipulation using only linear motion. Two parallel grippers hold different parts of an object while four translating fingertips move them relative to each other, enabling cap opening, pipetting, pumping, screwdriving, trigger actuation, and in-grasp reorientation across 35 objects, on both a robot arm and a humanoid.

We introduce the visible-reachable workspace (VRW), a design-stage measure that conditions visibility on feasible reaching configurations, and apply it to build Duke Humanoid V2, a 31-DoF humanoid with two independently actuated RGB-D cameras. Camera articulation raises visible-reachable coverage from 38% to 97%, and the dual-actuated design cuts two-target reach-and-grasp completion time by 17% and energy by 19% versus the same robot with fixed cameras.

We propose CEER, a compliant end-effector–root (EE-root) control abstraction for modular humanoid loco-manipulation within a hierarchical planning framework. CEER enables compliance-aware whole-body control in an interpretable task space defined by root motion commands and end-effector pose targets, and supports plug-and-play integration with heterogeneous high-level planners. Experiments in simulation and on hardware demonstrate 3.3 cm end-effector tracking accuracy with substantially reduced jerk, stable contact-rich manipulation under teleoperation, and up to 70% success in simulated single-object loco-manipulation tasks within a room-scale environment.

We introduce dynamic symmetry — the uniformity of a robot's attainable center-of-mass accelerations — and show that maximizing it consistently improves trajectory tracking, robustness, and energy efficiency across 1,000+ simulated morphologies. We build Argus, a 20-leg spherical robot achieving near-extreme dynamic isotropy, demonstrating orientation-invariant locomotion, agile terrain traversal, and resilience to partial actuator failures.

We present the Duke Humanoid, an open-source 10-DOF child-sized bipedal robot designed for energy-efficient locomotion using passive dynamics. We develop a reinforcement learning policy deployable zero-shot on hardware for velocity-tracking walking, and propose an end-to-end RL algorithm that encourages passive dynamics — reducing cost of transport by up to 50% in simulation and 31% in real-world tests.

Text2Robot converts user text descriptions into physical quadrupedal robots. Within minutes, text-to-3D models initialize diverse morphologies; within a day, geometric processing and body-control co-optimization produce a walking robot that accounts for real-world electronics and manufacturability. The framework enables rapid prototyping and opens new opportunities for generative robot design.

We present Flexipod, an open-source untethered quadrupedal soft robot platform for dynamic locomotion. The robot is 80 vol.% soft with 3D-printed gyroid-infill flexible legs that passively stabilize on multi-terrain environments. With gaits tuned in a CUDA-accelerated soft-body simulator, the real robot achieves 0.9 m/s (2.5 body lengths/sec) — faster than most untethered legged soft robots — and can execute backflips.