Motion Gen doesn't generate solution when enable_opt is set to TRUE and enable_graph is set to FALSE

[sanjkar126]

Hi,

I am using Motion Planning using curobo motion_gen. Unfortunately I am facing some issues with my configuration.

• Issue#1 : When I try to plan with obstacles motion_gen fails to find a solution even though there is one. I am able to confirm this because if I set enable_graph to TRUE and set enable_opt to FALSE then it finds a path where as if I change it the other way then it fails.

• Issue#2: With just enable_graph as TRUE; I can observe that the Path generated doesn't avoid the obstacles. This isn't consistent but I would say it happens 2 out of 10 times.

My configuration:

### Initialization ###
motion_gen_config = MotionGenConfig.load_from_robot_config(
            "my_robot.yml",
            WorldConfig(), # Empty World
            self.tensor_args,
            collision_checker_type=CollisionCheckerType.MESH,
            num_trajopt_seeds=12, 
            num_graph_seeds=12,
            use_cuda_graph=True,
            interpolation_dt=0.01, 
            maximum_trajectory_time=1.0,
            collision_cache={"obb": n_obstacle_cuboids, "mesh": n_obstacle_mesh},
            trajopt_dt=trajopt_dt,
            trajopt_tsteps=trajopt_tsteps,
            trim_steps=trim_steps,
            collision_activation_distance=0.0, # Not sure if this is needed
            interpolation_steps=35, # Had to set this or else it was generating 5000 Trajectory points
        )

 self.motion_gen = MotionGen(motion_gen_config)
 print("warming up...")
 self.motion_gen.warmup(enable_graph=False, warmup_js_trajopt=False)
  
### World Update ### (Obstacles are passed as ros2 Markers)
world_objs = WorldConfig.from_dict(
            {
                "cuboid": obstacle_dict
            }
        )

world_obj_mesh = world_objs.get_mesh_world()

world_cfg = WorldConfig(cuboid=world_objs.cuboid, mesh=world_obj_mesh)
self.motion_gen.clear_world_cache()
self.motion_gen.update_world(world_cfg)
  
 ### Planning ###       
 plan_config = MotionGenPlanConfig(
            enable_opt=False,
            enable_graph=True,
            #enable_graph_attempt=2,
            max_attempts=100,
 )
 result = self.motion_gen.plan_single(start_state, target_pose, plan_config)
        

[balakumar-s]

What did you set as values for these?
collision_cache={"obb": n_obstacle_cuboids, "mesh": n_obstacle_mesh},

[sanjkar126]

Tried setting self_collision_check=False and setting interpolation_steps=5000 with enable_opt=Ture & enable_graph=False. Still motion gen returns False.

[balakumar-s]

Can you also set max_trajectory_dt=2.0 to MotionGenConfig.load_from_robot_config() and see if that helps?

[sanjkar126]

Set maximum_trajectory_dt=2.0 but still No Solution Found.

motion_gen_config = MotionGenConfig.load_from_robot_config(
            "my_robot.yml",
            "collision_table.yml",
            self.tensor_args,
            collision_checker_type=CollisionCheckerType.PRIMITIVE,
            num_trajopt_seeds=2, # 12 is default
            num_graph_seeds=2,
            use_cuda_graph=True,
            interpolation_dt=0.01,
            maximum_trajectory_time=1.0,
            collision_cache={"obb": n_obstacle_cuboids, "mesh": n_obstacle_mesh},
            trajopt_dt=trajopt_dt,
            trajopt_tsteps=trajopt_tsteps,
            trim_steps=trim_steps,
            collision_activation_distance=0.0,
            maximum_trajectory_dt=2.0,
        )

[sanjkar126]

Bias node is not valid, not using bias node

Get this log line. Not sure if this indicates something

[balakumar-s]

That means that the robot is in collision at the retract_config. Try changing it so that there is no collision.

[Younch-15]

Current using v0.7.4 with Isaac Sim 4.0, facing exactly the same problem. When setting enable_graph as True, I always get FINETUNE_TRAJOPT_FAIL error no matter how I set enable_graph; when setting only enable_graph as True, the trajectory generated does not make any sense.