the match between given checkpoint and config

[ganyz]

Thanks for the great work!
There is a problem when I sim2real:
I use the the checkpoint below but the robot becomes unstable and moves on its own.

I want to check if the config matches the checkpoint, so I want to ask the config with the given checkpoint. Is the given g1_deepmimic_config.py?

Thanks so much!

[ArthurAllshire]

Is it unstable before switching to the heightmap policy? The procedure should be-- starting script, pressing start button, lowering it onto ground and then pressing button A to start the blind policy. Is it at least working up to this point?
If you are switching to the heightmap policy and it becomes unstable, try moving the robot around to build the heightmap before switching to that policy -- the LiDAR only sees like >1m from the robot so moving around to build it is needed.
One other thing to note is that for the heightmap policy, it does sometimes move on its own. This is because our data scale is not very large and so it's responsiveness to commands is not that great. But it should generally respond to what you do.

[ganyz]

Is it unstable before switching to the heightmap policy? The procedure should be-- starting script, pressing start button, lowering it onto ground and then pressing button A to start the blind policy. Is it at least working up to this point? If you are switching to the heightmap policy and it becomes unstable, try moving the robot around to build the heightmap before switching to that policy -- the LiDAR only sees like >1m from the robot so moving around to build it is needed. One other thing to note is that for the heightmap policy, it does sometimes move on its own. This is because our data scale is not very large and so it's responsiveness to commands is not that great. But it should generally respond to what you do.

Thanks for your helpful reply!

Now there is a video from our real experiment following your tips. The G1 robot tried several times to step onto the stairs but didn’t succeed at first. It finally managed to do it once ☺.

Is this behavior in the video normal ?, and are there more advice I should take care of to improve the performance ? ^.^

Thanks for your helpful reply again !

196dd9b705bde1e9ccac1d563710bfee.mp4

[ArthurAllshire]

I think this looks somewhat similar to when I was trying indoor experiments on the gantry. One issue is that elevation map doesnt work that well when on a gantry (as it has to reject suprious readings) and indoor when it hasn't moved around that much. Can you show what the heightmap looks like? As I mentioned above, you probably need to walk the robot around further from the stairs in order to build heightmap. hope it helps!

[ganyz]

I think this looks somewhat similar to when I was trying indoor experiments on the gantry. One issue is that elevation map doesnt work that well when on a gantry (as it has to reject suprious readings) and indoor when it hasn't moved around that much. Can you show what the heightmap looks like? As I mentioned above, you probably need to walk the robot around further from the stairs in order to build heightmap. hope it helps!

Thank you! Recently our team has less time to focus on this work, later we will follow your valuable advice to put into practice.

[Skvayzer]

I also noticed that the checkpoint was unstable and hard to control. The robot constantly rotates clockwise and goes in random direction. It applies to both real-world and simulation. Simulation is not supposed to have the problem with elevation map building since this map is simulated and always known. You can see the problem if you turn on the Manual Control mode and do not provide the control command. Please see the video attached (i tried every checkpoint and they all have this problem). This problem translates from simulation to the real-world deployment - it is veeery hard to control the robot.

Is this a problem with the config files or with the pretrained policy itself? @ArthurAllshire @hongsukchoi did you experience anything like that with the checkpoints that you shared?

Screencast.from.01-26-2026.02.25.35.PM.webm