Please inquire about how the pointnav_weights.pth file within WrappedPointNavResNetPolicy was trained.

[lovelyyoshino]

I have integrated VLFM into ROS, and during practical usage, I found that trained algorithms like pointnav_weights.pth can reach the target points, but they are unable to rotate 180 degrees in a short time, leading to inconsistencies in orientation. I would like to inquire about how to train my own navigation strategy.

[lovelyyoshino]

@jiuguangw @baxelrod-bdai I found that my target was behind me, so I simply stepped back, with only a slight tendency to turn.

[ros_objnav-1] Step: 48 | Target object: table
[ros_objnav-1] BLIP2ITMClient.cosine: (480, 640, 3), Seems like there is a table ahead., float cosine: 0.21706531941890717
[ros_objnav-1] Step: 48 | Target object: table
[ros_objnav-1] -----------------------Current goal: [-0.56651598 -0.40734302]
[ros_objnav-1] -----------------------current object: [-0.00011704 -0.00144525]
[ros_objnav-1] Navigating... tensor([[-0.0932, -0.9892]], device='cuda:0', grad_fn=)
[ros_objnav-1] Step: 48 | Mode: navigate | Action: [-0.09323746 -0.9891526 ]
[ros_objnav-1] ----------------------------action: tensor([[-0.0932, -0.9892]], device='cuda:0', grad_fn=)
[ros_objnav-1] Computed yaw: -0.041999642048971,x=-2.213651896454394e-05, y=-0.001332213170826435
[ros_objnav-1] Computed yaw: -0.041999642048971,x=-2.213651896454394e-05, y=-0.001332213170826435
[ros_objnav-1] Computed yaw: -0.041999642048971,x=-2.213651896454394e-05, y=-0.001332213170826435
[ros_objnav-1] Computed yaw: -0.041999642048971,x=-2.213651896454394e-05, y=-0.001332213170826435
[ros_objnav-1] Step: 49 | Target object: table
[ros_objnav-1] BLIP2ITMClient.cosine: (480, 640, 3), Seems like there is a table ahead., float cosine: 0.2091599702835083
[ros_objnav-1] Step: 49 | Target object: table
[ros_objnav-1] -----------------------Current goal: [-0.56651598 -0.40734302]
[ros_objnav-1] -----------------------current object: [ 3.33883276e-05 -1.46653466e-03]
[ros_objnav-1] Navigating... tensor([[0.7829, 0.8002]], device='cuda:0', grad_fn=)
[ros_objnav-1] Step: 49 | Mode: navigate | Action: [0.7828593 0.8002483]
[ros_objnav-1] ----------------------------action: tensor([[0.7829, 0.8002]], device='cuda:0', grad_fn=)
[ros_objnav-1] Computed yaw: -0.02648336092104131,x=-0.08454433083534241, y=0.0016461578197777271
[ros_objnav-1] Computed yaw: -0.02648336092104131,x=-0.08454433083534241, y=0.0016461578197777271
[ros_objnav-1] Computed yaw: -0.02648336092104131,x=-0.08454433083534241, y=0.0016461578197777271
[ros_objnav-1] Computed yaw: -0.02648336092104131,x=-0.08454433083534241, y=0.0016461578197777271
[ros_objnav-1] Step: 50 | Target object: table
[ros_objnav-1] BLIP2ITMClient.cosine: (480, 640, 3), Seems like there is a table ahead., float cosine: 0.20604312419891357
[ros_objnav-1] Step: 50 | Target object: table
[ros_objnav-1] -----------------------Current goal: [-0.56651598 -0.40734302]
[ros_objnav-1] -----------------------current object: [-0.0844821 0.0016965]
[ros_objnav-1] Navigating... tensor([[-0.2042, -0.9706]], device='cuda:0', grad_fn=)
[ros_objnav-1] Step: 50 | Mode: navigate | Action: [-0.20419289 -0.9706085 ]
[ros_objnav-1] ----------------------------action: tensor([[-0.2042, -0.9706]], device='cuda:0', grad_fn=)

[man-robot007]

hello could i contact you for some help ? i meet some question i canot solve. this is my wechat ywfyywq2024 thanks!