void displayPlannerData(const planning_scene::PlanningSceneConstPtr& planning_scene, const std::string &link_name) const { std::cout << "displayPlannerData called ... " << std::endl; const ompl_interface::ModelBasedPlanningContextPtr &pc = ompl_interface_->getLastPlanningContext(); if (pc) { ompl::base::PlannerData pd(pc->getOMPLSimpleSetup().getSpaceInformation()); pc->getOMPLSimpleSetup().getPlannerData(pd); robot_state::RobotState kstate = planning_scene->getCurrentState(); visualization_msgs::MarkerArray arr; std_msgs::ColorRGBA color; color.r = 1.0f; color.g = 0.25f; color.b = 1.0f; color.a = 1.0f; unsigned int nv = pd.numVertices(); for (unsigned int i = 0 ; i < nv ; ++i) { pc->getOMPLStateSpace()->copyToRobotState(kstate, pd.getVertex(i).getState()); kstate.getJointStateGroup(pc->getJointModelGroupName())->updateLinkTransforms(); const Eigen::Vector3d &pos = kstate.getLinkState(link_name)->getGlobalLinkTransform().translation(); visualization_msgs::Marker mk; mk.header.stamp = ros::Time::now(); mk.header.frame_id = planning_scene->getPlanningFrame(); mk.ns = "planner_data"; mk.id = i; mk.type = visualization_msgs::Marker::SPHERE; mk.action = visualization_msgs::Marker::ADD; mk.pose.position.x = pos.x(); mk.pose.position.y = pos.y(); mk.pose.position.z = pos.z(); mk.pose.orientation.w = 1.0; mk.scale.x = mk.scale.y = mk.scale.z = 0.025; mk.color = color; mk.lifetime = ros::Duration(30.0); arr.markers.push_back(mk); } pub_markers_.publish(arr); } }
bool ConvexConstraintSolver::solve(const planning_scene::PlanningSceneConstPtr& planning_scene, const moveit_msgs::GetMotionPlan::Request &req, moveit_msgs::GetMotionPlan::Response &res) const { // Need to add in joint limit constraints // Get the position constraints from the collision detector // The raw algorithm laid out in Chan et. al. is as follows: // 1. Compute unconstrained motion to go from proxy to goal. // 2. Get the current contact set by moving the proxy toward the goal by some epsilon and get all collision points. // 3. Compute constrained motion (convex solver). // 4. Compute collisions along the constrained motion path. // 5. Set proxy to stop at the first new contact along path. // In our framework this will look more like this: // Outside ths planner: // 1. Compute error between cartesian end effector and cartesian goal. // 2. Cap pose error (based on some heuristic?) because we are about to make a linear approximation. // 3. Use Jinverse to compute the joint deltas for the pose error (or perhaps we should frame this as a velocity problem). // 4. Add the joint deltas to the start state to get a goal state. // Inside the planner: // 1. Get the "current" contact set by moving the proxy toward the goal by some epsilon and get all collision points. // 2. Compute constrained motion (convex solver). // 3. Subdivide motion subject to some sort of minimum feature size. // 4. Move proxy in steps, checking for colliding state along the way. Optionally use interval bisection to refine. // - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -/ std::string group_name = req.motion_plan_request.group_name; std::string ee_group_name; std::string ee_control_frame; if(group_name == "right_arm"){ ee_group_name = "r_end_effector"; ee_control_frame = "r_wrist_roll_link"; } if(group_name == "left_arm") { ee_group_name = "l_end_effector"; ee_control_frame = "l_wrist_roll_link"; } ROS_WARN("Solving for group [%s], end-effector [%s], control frame [%s] using ConvexConstraintSolver!", group_name.c_str(), ee_group_name.c_str(), ee_control_frame.c_str()); // We "getCurrentState" just to populate the structure with auxiliary info, then copy in the transform info from the planning request. planning_models::KinematicState start_state = planning_scene->getCurrentState(); planning_models::robotStateToKinematicState(*(planning_scene->getTransforms()), req.motion_plan_request.start_state, start_state); // constrained_goal_state is the optimization output before interval stepping, proxy_state will be used as we step around. planning_models::KinematicState proxy_state = start_state; planning_models::KinematicState constrained_goal_state = start_state; std::string planning_frame = ros::names::resolve("/", planning_scene->getPlanningFrame()); const planning_models::KinematicState::JointStateGroup * jsg = proxy_state.getJointStateGroup(group_name); const planning_models::KinematicModel::JointModelGroup * jmg = planning_scene->getKinematicModel()->getJointModelGroup(group_name); // std::string end_effector_group_name = jmg->getAttachedEndEffectorGroupName(); // const std::vector<std::string>& subgroups = jmg->getSubgroupNames(); // ROS_INFO("End-effector group name is [%s].", end_effector_group_name.c_str()); // for(size_t i = 0; i< subgroups.size(); i++) // ROS_INFO("Subgroup [%zd] is [%s].", i, subgroups[i].c_str()); const planning_models::KinematicModel::JointModelGroup * ee_jmg = planning_scene->getKinematicModel()->getJointModelGroup(ee_group_name); // const std::vector<const planning_models::KinematicModel::JointModel*>& joint_models = jmg->getJointModels(); // std::map<std::string, unsigned int> joint_index_map; // for(size_t i = 0; i < joint_models.size(); i++) // { // //ROS_INFO("Group [%s] has joint %d: [%s]", req.motion_plan_request.group_name.c_str(), i, joint_models[i]->getName().c_str()); // joint_index_map[joint_models[i]->getName()] = i; // } const std::map<std::string, unsigned int>& joint_index_map = jmg->getJointVariablesIndexMap(); std::vector<double> limits_min, limits_max, joint_vector; std::vector<std::string> joint_names; limits_min.resize(7); limits_max.resize(7); joint_vector.resize(7); joint_names.resize(7); { const moveit_msgs::Constraints &c = req.motion_plan_request.goal_constraints[1]; for (std::size_t i = 0 ; i < c.joint_constraints.size() ; ++i) { std::string joint_name = c.joint_constraints[i].joint_name; if(joint_index_map.find(joint_name) == joint_index_map.end()) { ROS_WARN("Didin't find [%s] in the joint map, ignoring...", joint_name.c_str()); continue; } unsigned int joint_index = joint_index_map.find(joint_name)->second; std::vector<moveit_msgs::JointLimits> limits = planning_scene->getKinematicModel()->getJointModel(joint_name)->getLimits(); moveit_msgs::JointLimits limit = limits[0]; if(limit.has_position_limits) { limits_min[joint_index] = limit.min_position; limits_max[joint_index] = limit.max_position; } else { limits_min[joint_index] = -1E3; limits_max[joint_index] = 1E3; } joint_vector[joint_index] = start_state.getJointState(joint_name)->getVariableValues()[0]; joint_names[joint_index] = joint_name; //ROS_INFO("Joint [%d] [%s] has min %.2f, value %.2f, max %.2f", joint_index, joint_name.c_str(), limits_min[joint_index], joint_vector[joint_index], limits_max[joint_index]); } } // ======== Extract all contact points and normals from previous collision state, get associated Jacobians ======== std::vector<Eigen::MatrixXd> contact_jacobians; std::vector<Eigen::Vector3d> contact_normals; for( collision_detection::CollisionResult::ContactMap::const_iterator it = last_collision_result.contacts.begin(); it != last_collision_result.contacts.end(); ++it) { std::string contact1 = it->first.first; std::string contact2 = it->first.second; const std::vector<collision_detection::Contact>& vec = it->second; for(size_t contact_index = 0; contact_index < vec.size(); contact_index++) { Eigen::Vector3d point = vec[contact_index].pos; Eigen::Vector3d normal = vec[contact_index].normal; double depth = vec[contact_index].depth; ROS_INFO("Contact between [%s] and [%s] point: %.2f %.2f %.2f normal: %.2f %.2f %.2f depth: %.3f", contact1.c_str(), contact2.c_str(), point(0), point(1), point(2), normal(0), normal(1), normal(2), depth); // Contact point needs to be expressed with respect to the link; normals should stay in the common frame std::string group_contact; if( jmg->hasLinkModel(contact1) || ee_jmg->hasLinkModel(contact1) ) group_contact = contact1; else if( jmg->hasLinkModel(contact2) || ee_jmg->hasLinkModel(contact2) ) group_contact = contact2; else { ROS_WARN("Contact isn't on group [%s], skipping...", req.motion_plan_request.group_name.c_str()); continue; } planning_models::KinematicState::LinkState *link_state = start_state.getLinkState(group_contact); Eigen::Affine3d link_T_world = link_state->getGlobalCollisionBodyTransform().inverse(); point = link_T_world*point; Eigen::MatrixXd jacobian; if(jsg->getJacobian(group_contact, point, jacobian)) { contact_jacobians.push_back(jacobian); contact_normals.push_back(normal); } } } // ======== Extract goal "constraints" ======== const moveit_msgs::Constraints &c = req.motion_plan_request.goal_constraints[0]; // Position and Orientation if(c.position_constraints.size() != 1 || c.orientation_constraints.size() != 1) { ROS_ERROR("Currently require exactly one position and orientation constraint. Aborting..."); return false; } moveit_msgs::PositionConstraint pc = c.position_constraints[0]; moveit_msgs::OrientationConstraint oc = c.orientation_constraints[0]; pc.header.frame_id = ros::names::resolve("/", pc.header.frame_id); oc.header.frame_id = ros::names::resolve("/", oc.header.frame_id); if(pc.link_name != oc.link_name) { ROS_ERROR("Currently can't support position and orientation goals that are not for the same link. Aborting..."); return false; } if(pc.constraint_region.primitive_poses.size() == 0) { ROS_ERROR("Need to specify a single primitive_pose for position constraint region. Aborting..."); return false; } if(pc.constraint_region.primitive_poses.size() != 1) { ROS_ERROR("Need exactly one 'pose' for the end-effector goal region. Aborting..."); } if(pc.header.frame_id != planning_frame) ROS_WARN("The position goal header [%s] and planning_frame [%s] don't match, things are probably all wrong!", pc.header.frame_id.c_str(), planning_frame.c_str() ); if(oc.header.frame_id != planning_frame) ROS_WARN("The orientation goal header [%s] and planning_frame [%s] don't match, things are probably all wrong!", oc.header.frame_id.c_str(), planning_frame.c_str() ); Eigen::Vector3d goal_point; Eigen::Quaterniond goal_quaternion_e; { // scoped so we don't pollute function scope with these message temps geometry_msgs::Point msg_goal_point = pc.constraint_region.primitive_poses[0].position; geometry_msgs::Quaternion msg_goal_orientation = oc.orientation; goal_point = Eigen::Vector3d(msg_goal_point.x, msg_goal_point.y, msg_goal_point.z); goal_quaternion_e = Eigen::Quaterniond(msg_goal_orientation.w, msg_goal_orientation.x, msg_goal_orientation.y, msg_goal_orientation.z); } //ROS_INFO("Computing position and orientation error..."); planning_models::KinematicState::LinkState *link_state = start_state.getLinkState(pc.link_name); Eigen::Affine3d planning_T_link = link_state->getGlobalLinkTransform(); Eigen::Vector3d ee_point_in_ee_frame = Eigen::Vector3d(pc.target_point_offset.x, pc.target_point_offset.y, pc.target_point_offset.z); Eigen::Vector3d ee_point_in_planning_frame = planning_T_link*ee_point_in_ee_frame; // TODO need to make sure these are expressed in the same frame. Eigen::Vector3d x_error = goal_point - ee_point_in_planning_frame; Eigen::Vector3d delta_x = x_error; double x_error_mag = x_error.norm(); double LINEAR_CLIP = 0.02; if(x_error_mag > LINEAR_CLIP) delta_x = x_error/x_error_mag*LINEAR_CLIP; // = = = = Rotations are gross = = = = Eigen::Quaterniond link_quaternion_e = Eigen::Quaterniond(planning_T_link.rotation()); tf::Quaternion link_quaternion_tf, goal_quaternion_tf; tf::RotationEigenToTF( link_quaternion_e, link_quaternion_tf ); tf::RotationEigenToTF( goal_quaternion_e, goal_quaternion_tf ); tf::Quaternion delta_quaternion = link_quaternion_tf.inverse()*goal_quaternion_tf; double rotation_angle = delta_quaternion.getAngle(); //tf::Vector3 rotation_axis = delta_quaternion.getAxis(); //ROS_INFO("Delta is [%.3f] radians about [%.3f, %.3f, %.3f]", rotation_angle, rotation_axis.x(), rotation_axis.y(), rotation_axis.z()); double ANGLE_CLIP = 0.2; double clipped_rotation_fraction = std::min<double>(1.0, ANGLE_CLIP/fabs(rotation_angle)); if(clipped_rotation_fraction < 0) ROS_ERROR("Clipped rotation fraction < 0, look into this!"); tf::Matrix3x3 clipped_delta_matrix; tf::Quaternion clipped_goal_quaternion = link_quaternion_tf.slerp(goal_quaternion_tf, clipped_rotation_fraction); clipped_delta_matrix.setRotation( link_quaternion_tf.inverse()*clipped_goal_quaternion ); tf::Vector3 delta_euler; clipped_delta_matrix.getRPY(delta_euler[0], delta_euler[1], delta_euler[2]); //ROS_INFO("Delta euler in link frame = [%.3f, %.3f, %.3f]", delta_euler[0], delta_euler[1], delta_euler[2]); tf::Matrix3x3 link_matrix(link_quaternion_tf); delta_euler = link_matrix * delta_euler; //ROS_INFO("Delta euler in planning_frame = [%.3f, %.3f, %.3f]", delta_euler[0], delta_euler[1], delta_euler[2]); // Get end-effector Jacobian // std::string ee_control_frame = "r_wrist_roll_link"; // if(false && jmg->isChain()) // ee_control_frame = planning_scene->getKinematicModel()->getJointModelGroup(jmg->getAttachedEndEffectorGroupName())->getEndEffectorParentGroup().second; // else // ROS_WARN("Using r_wrist_roll_link as HARD_CODED value."); if(ee_control_frame != pc.link_name) ROS_WARN("ee_control_frame [%s] and position_goal link [%s] aren't the same, this could be bad!", ee_control_frame.c_str(), pc.link_name.c_str()); // ROS_INFO("Getting end-effector Jacobian for local point %.3f, %.3f, %.3f on link [%s]", // ee_point_in_ee_frame(0), // ee_point_in_ee_frame(1), // ee_point_in_ee_frame(2), // ee_control_frame.c_str()); Eigen::MatrixXd ee_jacobian; if(!jsg->getJacobian(ee_control_frame , ee_point_in_ee_frame , ee_jacobian)) { ROS_ERROR("Unable to get end-effector Jacobian! Can't plan, exiting..."); return false; } //ROS_INFO_STREAM("End-effector jacobian in planning frame is: \n" << ee_jacobian); // ======== Pack into solver data structure, run solver. ======== //ROS_INFO("Packing data into the cvx solver..."); // Vars vars; // Params params; // Workspace work; // Settings settings; // CVX Settings cvx.set_defaults(); cvx.setup_indexing(); cvx.settings.verbose = 0; // - - - - - - - load all problem instance data - - - - - - - // unsigned int N = 7; // number of joints in the chain // end-effector Jacobian // TODO magic numbers (though I suppose the CVX solver is already hard-coded) for(unsigned int row = 0; row < 3; row++ ) { for(unsigned int col = 0; col < N; col++ ) { // CVX matrices are COLUMN-MAJOR!!! cvx.params.J_v[col*3 + row] = ee_jacobian(row, col); cvx.params.J_w[col*3 + row] = ee_jacobian(row+3, col); } } // Set weights for objective terms cvx.params.weight_x[0] = 1.0; // translational error cvx.params.weight_w[0] = 0.1; // angular error (error in radians is numerically much larger than error in meters) cvx.params.weight_q[0] = 0.001; // only want to barely encourage values to stay small... // set up constraints from contact set unsigned int MAX_CONSTRAINTS = 25; unsigned int constraint_count = std::min<size_t>(MAX_CONSTRAINTS, contact_normals.size()); for(unsigned int constraint = 0; constraint < MAX_CONSTRAINTS; constraint++) { if(constraint < constraint_count) { // CVX matrices are COLUMN-MAJOR!!! for (int j = 0; j < 3*7; j++) { cvx.params.J_c[constraint][j] = contact_jacobians[constraint](j%3, j/3); } for (int j = 0; j < 3; j++) { cvx.params.normal[constraint][j] = contact_normals[constraint][j]; } } else{ //printf("setting to zero\n"); for (int j = 0; j < 3*7; j++) { cvx.params.J_c[constraint][j] = 0; } for (int j = 0; j < 3; j++) { cvx.params.normal[constraint][j] = 0; } } } for(unsigned int index = 0; index < N; index++ ) { cvx.params.q[index] = joint_vector[index]; cvx.params.q_min[index] = limits_min[index]; cvx.params.q_max[index] = limits_max[index]; } // TODO should these be clipped down at all? :) for(unsigned int index = 0; index < 3; index++ ) { cvx.params.x_d[index] = delta_x(index); // TODO magic minus sign!! cvx.params.w_d[index] = delta_euler[index]; } // - - - - - - - Solve our problem at high speed! - - - - - - - // long num_iters = 0; num_iters = cvx.solve(); if(!cvx.work.converged) { printf("solving failed to converge in %ld iterations.\n", num_iters); return false; } // ======== Unpack solver result into constrained goal state. ======== ROS_INFO("Finished solving in %ld iterations, unpacking data...", num_iters); Eigen::VectorXd joint_deltas(N); std::map<std::string, double> goal_update; for(size_t joint_index = 0; joint_index < joint_names.size(); joint_index++) { std::string joint_name = joint_names[joint_index]; joint_deltas(joint_index) = cvx.vars.q_d[joint_index]; goal_update[joint_name] = cvx.vars.q_d[joint_index] + start_state.getJointState(joint_name)->getVariableValues()[0]; //ROS_INFO("Updated joint [%zd] [%s]: %.3f + %.3f", joint_index, joint_name.c_str(), start_state.getJointState(joint_name)->getVariableValues()[0], cvx.vars.q_d[joint_index]); } Eigen::VectorXd cartesian_deltas = ee_jacobian*joint_deltas; //ROS_INFO("Raw Error: translate [%.3f, %.3f, %.3f]", // x_error(0), x_error(1), x_error(2)); ROS_INFO("ClipError: translate [%.3f, %.3f, %.3f] euler [%.2f, %.2f, %.2f]", delta_x(0), delta_x(1), delta_x(2), delta_euler[0], delta_euler[1], delta_euler[2]); ROS_INFO("Output: translate [%.3f, %.3f, %.3f] euler [%.2f, %.2f, %.2f]", cartesian_deltas(0), cartesian_deltas(1), cartesian_deltas(2), cartesian_deltas(3), cartesian_deltas(4), cartesian_deltas(5)); //goal_update[vars.qdd_c[index] + ] constrained_goal_state.setStateValues(goal_update); // ======== Step toward constrained goal, checking for new collisions along the way ======== //double proxy_goal_tolerance = 0.1; double interpolation_progress = 0.0; double interpolation_step = 0.34; collision_detection::CollisionResult collision_result; while(interpolation_progress <= 1.0) { //ROS_INFO("Doing interpolation, with progress %.2f ", interpolation_progress); // TODO this interpolation scheme might not allow the arm to slide along contacts very well... planning_models::KinematicStatePtr point; point.reset(new planning_models::KinematicState(constrained_goal_state)); start_state.interpolate(constrained_goal_state, interpolation_progress, *point); // get contact set collision_detection::CollisionRequest collision_request; // TODO magic number collision_request.max_contacts = 50; collision_request.contacts = true; collision_request.distance = false; collision_request.verbose = false; collision_result.clear(); planning_scene->checkCollision(collision_request, collision_result, *point); if(collision_result.collision) { break; } else { proxy_state = *point; } // TODO Use proxy_goal_tolerance to exit if we are close enough! interpolation_progress += interpolation_step; } // Store the last collision result! last_collision_result = collision_result; // ======== Convert proxy state to a "trajectory" ======== //ROS_INFO("Converting proxy to a trajectory, and returning..."); moveit_msgs::RobotTrajectory rt; trajectory_msgs::JointTrajectory traj; trajectory_msgs::JointTrajectoryPoint pt; sensor_msgs::JointState js; planning_models::kinematicStateToJointState(proxy_state, js); //const planning_models::KinematicModel::JointModelGroup *jmg = planning_scene->getKinematicModel()->getJointModelGroup(req.motion_plan_request.group_name); // getJointNames for(size_t i = 0 ; i < js.name.size(); i++) { std::string name = js.name[i]; if( !jmg->hasJointModel(name) ) continue; traj.joint_names.push_back(js.name[i]); pt.positions.push_back(js.position[i]); if(js.velocity.size()) pt.velocities.push_back(js.velocity[i]); } pt.time_from_start = ros::Duration(req.motion_plan_request.allowed_planning_time*1.0); traj.points.push_back(pt); traj.header.stamp = ros::Time::now(); traj.header.frame_id = "odom_combined"; res.trajectory.joint_trajectory = traj; return true; }