Bingjie Tang

We present Refinery, a framework for improving contact-rich manipulation policies through active fine-tuning and deployment-time optimization. The method identifies informative real-world interactions, adapts policies after deployment, and optimizes behavior online, helping robots maintain robust performance under contact uncertainty and task variation.

We present FORGE, a method for robust sim-to-real transfer in contact-rich manipulation under pose uncertainty. During simulation training, FORGE combines force-guided exploration with dynamics randomization so policies can tolerate misalignment and uncertainty, then transfer reliably to real robotic assembly and manipulation settings.

We introduce SRSA, a framework for solving new robotic assembly tasks by retrieving and adapting skills from a pre-existing policy library. The method identifies relevant prior skills, transfers them to new geometries and task conditions, and reduces the need to learn every assembly behavior from scratch.

We propose MatchMaker, a pipeline for automatically generating diverse, simulation-compatible assembly asset pairs. By creating geometrically varied part combinations at scale, MatchMaker reduces the manual effort required to build assembly benchmarks and provides richer training assets for learning robust robotic assembly skills.

We present a learning framework for assembly across many geometries, including a dataset of 100 simulation- and real-world-compatible assemblies. AutoMate learns specialist and generalist policies in parallel simulation, solves many assembly families with high success rates, and transfers zero-shot to real robots for end-to-end assembly.

We introduce algorithms and tools for transferring contact-rich assembly policies from simulation to the real world. The system combines simulation-aware policy updates, signed-distance-field rewards, sampling-based curricula, and a policy-level action integrator, enabling robots to solve pick, place, and insertion tasks with repeatable real-world performance.

We develop an image-based method for selective tabletop rearrangement in clutter with a parallel-jaw gripper. The system chooses which object to move, decides whether and where to push or grasp, and uses visual correspondence to place grasped objects, producing targeted rearrangements from raw visual observations.

We train self-supervised policies that combine planar pushing with 6-degree-of-freedom grasping for dense tabletop clutter. By coordinating pushes with a richer grasp action space, the robot can rearrange obstacles and recover feasible grasps, enabling more flexible manipulation across diverse household objects and cluttered scenes.