Manipulation Motion Planning

Isaac SDK provides the following components for motion planning of robotics arms:

  • EndEffectorGlobalPlanner: Converts cartesian target for the end effector to joint angles target using inverse kinematics. This codelet can either receive the cartesian target (3d pose) as composite message, or read from the pose tree.

  • MultiJointPlanner: Produces a trajectory to move a list of joints from a starting state to a target state. This component contains generic parsing code for the messages, and requires an additional component that implements the MultiJointPlannerInterface.

  • MultiJointLqrPlanner: Implements MultiJointPlannerInterface using LQR. The LQR solver treats each joint is treated independently, and automatically adjusts the time to find a valid trajectory that does not exceed the minimum and maximum speed and acceleration constraints. It also ensures that the last state in the trajectory is the received target.

    This component only outputs a new plan when the target has changed. Use the speed_max, speed_min, acceleration_max, and acceleration_min config parameters to ensure the plan is feasible for the robot.

  • MultiJointRmpPlanner: Implements MultiJointPlannerInterface using RMPFlow. It outputs short trajectory by integrating the output of the RMPFlow policy over a number of timesteps. Because RMPFlow acts as a local reactive controller, it expects the target state to not be too far away from the starting position. See the paper below for more information on how RMPFlow works.

  • MultiJointController: Interpolates a trajectory received from the planner to determine the current command to send. This codelet can output either a joint position or speed command based on the control_mode config. The interpolation is based on the timestamps in the trajectory and the current tick time–with a small look-ahead, which can be configured with the command_delay config parameter, to account for latency in the control loop.

These codelets require a kinematic_tree config parameter, which must refer to a kinematic-tree file. The codelets use the kinematic-tree configuration to retrieve the number of joints and their names for parsing and serializing CompositeProto messages. See the Manipulation Kinematics documentation for more information on creating a kinematic-tree file.

Isaac SDK provides a subgraph that integrates the MultiJointPlanner, MultiJointController and KinematicTree components. You can find this subgraph at packages/planner/apps/multi_joint_lqr_control.subgraph.json (using MultiJointLqrPlanner) or packages/planner/apps/multi_joint_rmp_control.subgraph.json (using MultiJointRmpPlanner). The subgraphs provide the following interface edges:

  • state (input): “subgraph/interface/joint_state”
  • target (input): “subgraph/interface/joint_target”
  • command (output): “subgraph/interface/joint_command”

Manipulation-related components use CompositeProto messages to communicate. A CompositeProto with rank-1 tensor is used for a single state or command. The following example shows a state for two joints:


CompositeProto: { "schema": [ {"entity": "shoulder", "element_type": Float64, "measure": Position}, {"entity": "shoulder", "element_type": Float64, "measure": Speed}, {"entity": "elbow", "element_type": Float64, "measure": Position}, {"entity": "elbow", "element_type": Float64, "measure": Speed} ], "schema_hash": "...", "values": { "element_type": Float64, "sizes": [19], "dataBufferIndex": 0 } } buffers: [[0.7, 0.3, -0.4, 0.1]]

A CompositeProto with rank-2 tensor represents a trajectory. The first dimension of the tensor is the number of timesteps. The timestamp for each timestep is part of the schema. The following example shows a trajectory for two joints with two timesteps at t=0 and t=0.1:


CompositeProto: { "schema": [ {"entity": "timestamp", "element_type": Float64, "measure": Time}, {"entity": "joint1", "element_type": Float64, "measure": Position}, {"entity": "joint2", "element_type": Float64, "measure": Position}, {"entity": "joint1", "element_type": Float64, "measure": Speed}, {"entity": "joint2", "element_type": Float64, "measure": Speed} ], "schema_hash": "...", "values": { "element_type": Float64, "sizes": [10 , 4], "dataBufferIndex": 0 } } buffers: [[[0, 0.3, 0.7, -0.4, 0.1], [0.1, 0.32, 0.72, -0.3, 0.15]]]

Previous Manipulation
Next Manipulation Kinematics
© Copyright 2018-2020, NVIDIA Corporation. Last updated on Apr 19, 2024.