Systems and techniques for grasp selection may include receiving one or more candidate object trajectories and a current grasp of a robotic hand on an object, sampling random candidate grasps for the one or more candidate object trajectories based on the current grasp, generating one or more grasps to be optimized for each of the one or more candidate object trajectories based on the sampled candidate grasps, and optimizing one or more of the grasps to be optimized for each of the one or more candidate object trajectories based on a cost function.

Systems and methods for online iterative re-planning are provided herein. In one embodiment, a method includes receiving, at a first time step, a first grasp and an initial object pose of an agent. The method also includes generating a first set of candidate object trajectories based on the first grasp and the initial object pose. Candidate object trajectories of the first set of candidate object trajectories provide a number object poses from the initial object pose to a goal for a number of future time steps after the first time step. The method further includes calculating contact points for grasps associated with each candidate object trajectory of the first set of candidate object trajectories. The method further includes selecting a first candidate object trajectory from the first set of candidate object trajectories. The method includes causing the agent to execute the first candidate object trajectory at a second time step.

A system for generating a model-free reinforcement learning policy for a robotic hand for grasping an object is provided, including a processor; a memory; and a simulator implemented via the processor and the memory, performing: sampling a plurality of stable grasps relevant to reorienting the grasped object about a desired axis of rotation and using stable grasps as initial states for collecting training trajectories; learning finger-gaiting and finger-grasping policies for each axis of rotation in the hand coordinate frame based on proprioceptive sensing in the robotic hand, wherein the finger-gaiting and finger-pivoting policy is implemented on the robotic hand.

Systems and methods for online iterative re-planning are provided herein. In one embodiment, a method includes receiving, at a first time step, a first grasp and an initial object pose of an agent. The method also includes generating a first set of candidate object trajectories based on the first grasp and the initial object pose. Candidate object trajectories of the first set of candidate object trajectories provide a number object poses from the initial object pose to a goal for a number of future time steps after the first time step. The method further includes calculating contact points for grasps associated with each candidate object trajectory of the first set of candidate object trajectories. The method further includes selecting a first candidate object trajectory from the first set of candidate object trajectories. The method includes causing the agent to execute the first candidate object trajectory at a second time step.

Systems and techniques for grasp selection may include receiving one or more candidate object trajectories and a current grasp of a robotic hand on an object, sampling random candidate grasps for the one or more candidate object trajectories based on the current grasp, generating one or more grasps to be optimized for each of the one or more candidate object trajectories based on the sampled candidate grasps, and optimizing one or more of the grasps to be optimized for each of the one or more candidate object trajectories based on a cost function.