FLASH: Fast Learning via GPU-Accelerated Simulation for High-Fidelity Deformable Manipulation in Minutes

Figure 2 : An illustration of the FLASH System. FLASH is a GPU-accelerated high-fidelity deformable simulation framework that enables fast, large-scale learning for robotic manipulation. On top of FLASH, we build a learning and control pipeline that constructs observations from proprioception and perception, trains deployable policies via imitation learning, and transfers the learned policy to real robots for deformable manipulation tasks.

Figure 8 : Snapshots from the bimanual shirt folding example, simulated using our FLASH simulator. The dual end-effectors execute a sequential folding strategy: the left gripper first manipulates the left sleeve to the center and releases it, followed by the right gripper completing the fold for the opposing sleeve. The end-effector trajectories are generated via kinematic interpolation through a sequence of pre-defined heuristic waypoints.

Figure 12 : Real-Sim Observation Alignment and Depth Augmentation. Left: The physical hardware setup featuring an overhead RGB-D camera and the resulting segmented RGB observation . Note that the real observation (bottom left) presents obvious robot arm occlusions. Right: Our simulation rendering pipeline designed to bridge this domain gap. We sequentially apply random blockout to mimic segmentation imperfections, boundary jitter & noise to match sensor characteristics, and robot self-occlusion to reproduce the physical setup. The final composite augmentation (bottom right) serves as the robus

Figure 14 : Student policy architecture. Our student policy outputs actions and state reconstruction from proprioception and perception observations.

Figure 5 : Learned policies in simulation. (a) A single arm folds the towel. (b) A humanoid folds the towel. (c) Two arms fold the towel together. (d) Two arms fold the T-shirt. (e) Two arms fold the shorts.

Figure 6 : Robustness to external disturbances. (a.1–a.3) Recovery from unexpected displacement by human. (b.1–b.3) Recovery from the towel being pulled away . The system maintains task continuity without human intervention.

Figure 3 : Single environment simulation results. We performed simulation experiments within a single environment, (a) Friction validation demonstrating the simulator’s ability to distinguish coefficients with 0.001 precision. (b) Cloth simulation fidelity tested via cloth-chair collision. (c) A soft gripper grasping a soft cube via friction.

Figure 4 : T-shirt dual-sleeve folding results. We compare the final results of different simulators and parameters after folding a T-shirt with a fixed trajectory. (a) Comparison with real world and baselines. We compare against representative GPU-capable simulators: Genesis (PBD), Isaac Sim (FEM), Newton (VBD), and real execution; FLASH most closely matches real behavior, producing smooth, symmetric folds that settle stably with accurate frictional sticking, while baselines exhibit various failure features. (b) Simulation parameter ablation (FLASH). With the trajectory fixed (Sec. V.A), vary

Figure 7 : Qualitative real-world results. (Left) Keyframe visualization of bimanual folding ( t_{1}\to t_{3} ) for (a) shorts and (b) T-shirt. (Right) Snapshots from the continuous evaluation on AdamU at 20, 40, and 60 min (c.1–c.3). Main frames show egocentric camera observations; insets show the third-person view and clock time.

Figure 10 : Visualization of Extrinsic Alignment and Scene Calibration. (a.1) & (b.1) Real-world Validation: The accumulated point clouds for Adam-U (a.1) and the sub-pixel reprojection errors ( <0.05 px) for Airbot (b.1) confirm high-precision geometric alignment. (a.2) & (b.2) Simulation Alignment: The corresponding digital twin environments are shown on the right. In both setups, the calibrated camera frame is visualized overhead, where the blue arrow explicitly denotes the Z-axis aligned with the optical axis.