00:38:37 James Newton: Safety “Suggestions”
00:40:20 James Newton: I watch our Roombas get high centered and then recover, and not lose position quite often.
00:41:37 Adéna: Hi everyone, I’m new here,thought I’d see what the group is about.
00:42:16 Adéna: I’m working on automated robotic learning. Anyone here working on something similar?
00:42:20 tjame: Welcome! Happy to have all new comers.
00:42:47 tjame: Definitely, all the robots built for the SRS competitions have to be autonomous.
00:43:16 tjame: Although you may be referring to AI, in which case that is more rare.
00:44:37 Adéna: Indeed
00:44:48 Adéna: Indeed I refer to AI.
00:45:16 James Newton: “automated robotic learning” is a pretty wide net. Can you share a bit more about your specific focus?
00:51:15 Adéna: Like this: https://leap-workshop.github.io/index.html#schedule
00:54:43 James Newton: I’ve done a tiny bit of work with old school AI, e.g. PDDLs used to do planning. And with small NN’s to learn kinematics and limited jerk motion plans. I’m not sure anyone in this group has done any large scale LLM type robotic control. I’ve tried once but found the hallucination issue was killer.
00:56:21 Keith: Was it this one ? https://www.youtube.com/watch?v=xLO2r0fhZZg
00:56:57 Adéna: @JN – do you refer to generative LLM hallucinations?
00:58:15 tjame: Not that robot, but I appreciate the video.
00:58:47 Craig Austin: https://www.youtube.com/watch?v=ZQ4Sqihq_kQ&pp=ygUdcm9ib3QgY2hhbmdlcyBpdHMgb3duIGJhdHRlcnk%3D
01:00:19 Craig Austin: Robot that changes its own battery
01:04:27 tjame: Spot with Fluke’s SV600 acoustic imager array for plant inspection:
01:04:31 James Newton: Reacted to “@JN – do you refer…” with 👍
01:19:03 James Newton: What is a px4.ulog?
01:19:52 Shouheng: That PX4 ULogs are logs from a PX4 drone flight.
01:20:06 James Newton: PX4 is a model of drone?
01:20:10 Shouheng: https://docs.px4.io/main/en/dev_log/ulog_file_format
01:20:21 James Newton: Reacted to “https://docs.px4.i…” with 👍
01:20:57 James Newton: How trustworthy are those numerical calculations? LLMs often hallucinate math.
01:21:41 Shouheng: Reacted to “How trustworthy ar…” with ❤️
01:21:49 Shouheng: Arun is covering this topic right now 🙂
01:21:53 James Newton: Reacted to “Arun is covering t…” with 👍
01:22:41 James Newton: Smart! use the LLM to code up HOW to do the calculation and then let software for calculation actually execute it.
01:22:53 Shouheng: Exactly
01:23:47 James Newton: I wonder if the name brand AI’s are starting to do that (e.g. ChatGPT, Gemni, etc…) There was talk early on of building LLM interfaces to Wolfram Alpha, etc…
01:30:37 Connor: If asked something like, “give me a timeline of events”, would it be able to detect a crash, takeoff, failure, …etc? Or would that just be dependent on the quality of model used?
01:32:25 Connor: Understood. Initial suspicion goes a long way.
01:36:07 Shouheng: Apache Arrow
01:37:09 tjame: How many data runs are required to get accurate answers?
01:52:46 James Newton: Craig, I’ve actually done something like that with Gemini. If you can feed it labeled data (e.g. “when I sent this value to this address, I got this datastream”) and several examples of that. I asked it to write a program to decode it, and it got… far enough there that I was able to finish it off.
01:55:04 Shouheng: https://github.com/Extelligence-ai/bagel
01:55:08 tjame: Links to their discord server is on GitHub.
01:58:42 James Newton: For an example of a PDDL that both develops and checks a plan, and then executes it in an online simulator see
https://github.com/lmarsbrown/STRIPS/blob/master/examples/stacking_blocks/BlocksDexter.dde