Finn Busch

PhD Student in Robotics/Computer Science at KTH Royal Institute of Technology

[Finn Busch @fnnBsch]

Teaching a robot to navigate should be as simple as recording a video with your phone.

In LoTIS, we study how an RGB video can guide a robot without camera calibration, poses, or robot-specific training.

Page, code, demo:
https://t.co/r82gOIYu2i

#RSS2026 #ICRA2026

1/9 🧵

[Finn Busch @fnnBsch]

@JasonJZLiu I see, thanks :) I assume the same but I guess it might not always be true that smaller join torque errors always mean smaller EE errors? Either way, nice work!

[Finn Busch @fnnBsch]

@JasonJZLiu Transformed to end effector forces

[Finn Busch @fnnBsch]

Reach out if you want to chat and are at #ICRA2026, or if you want to see a live demo of our method with the @Bitcraze_se drones :)

[Finn Busch @fnnBsch]

Teaching a robot to navigate should be as simple as recording a video with your phone.

In LoTIS, we study how an RGB video can guide a robot without camera calibration, poses, or robot-specific training.

Page, code, demo:
https://t.co/r82gOIYu2i

#RSS2026 #ICRA2026

1/9 🧵

[Finn Busch @fnnBsch]

@giffmana The hard part might be getting the ghost registered in the right place in the view,

Reminds me of some work we did recently on localizing a previous recording directly in the current view, maybe there’s some overlap: https://t.co/9jVXRWjp6A

[Finn Busch @fnnBsch]

Teaching a robot to navigate should be as simple as recording a video with your phone.

In LoTIS, we study how an RGB video can guide a robot without camera calibration, poses, or robot-specific training.

Page, code, demo:
https://t.co/r82gOIYu2i

#RSS2026 #ICRA2026

1/9 🧵

[Finn Busch @fnnBsch]

@TobiasFischer11 Thanks for the insights :) might try to apply this to a work of ours that is related (https://t.co/9jVXRWjp6A). Was it easy to get this to train stably?

[Finn Busch @fnnBsch]

Teaching a robot to navigate should be as simple as recording a video with your phone.

In LoTIS, we study how an RGB video can guide a robot without camera calibration, poses, or robot-specific training.

Page, code, demo:
https://t.co/r82gOIYu2i

#RSS2026 #ICRA2026

1/9 🧵

[Finn Busch @fnnBsch]

@TobiasFischer11 Interesting! Given that, and the findings in tab 4 which suggest that weight sharing gives better perf, do you think training your model with a larger encoder would get even better performance? I.e. same compute as prior methods?

[Finn Busch @fnnBsch]

Try the model yourself on your own data at https://t.co/8UAS757bpy and let us know what you think :)

Find us at #RSS2026 in Sydney in July or the MM-SpatialAI workshop @ #ICRA2026 in Vienna next week!

Thanks to my amazing colleagues for making this project possible :)

9/9

[Finn Busch @fnnBsch]

Teaching a robot to navigate should be as simple as recording a video with your phone.

In LoTIS, we study how an RGB video can guide a robot without camera calibration, poses, or robot-specific training.

Page, code, demo:
https://t.co/r82gOIYu2i

#RSS2026 #ICRA2026

1/9 🧵

[Finn Busch @fnnBsch]

Results

The largest gains appear when the setting becomes less like “follow the video as recorded”:

off-traj starts, camera mismatch, and especially backward traversal.

In sim and real-world experiments, LoTIS substantially outperforms prior baselines.

8/9

[Finn Busch @fnnBsch]

@iamgrigorev Same here for vision tasks. Notice a small hit on training throughput, would you share the kernel?

[Finn Busch @fnnBsch]

@maharshii Interesting, this is related to your earlier tweet right? Does this also mean no matter if it is dynamic=True or False, the int/float Inputs will not be compiled into the kernel and you can change them at runtime?

[Finn Busch @fnnBsch]

@maharshii Please let me know if you figure this out :)