Our LLM-controlled office robot can't pass butter (andonlabs.com)

229 points by lukaspetersson ↗ HN
Hi HN! Our startup, Andon Labs, evaluates AI in the real world to measure capabilities and to see what can go wrong. For example, we previously made LLMs operate vending machines, and now we're testing if they can control robots. There are two parts to this test:

1. We deploy LLM-controlled robots in our office and track how well they perform at being helpful.

2. We systematically test the robots on tasks in our office. We benchmark different LLMs against each other. You can read our paper "Butter-Bench" on arXiv: https://arxiv.org/pdf/2510.21860

The link in the title above (https://andonlabs.com/evals/butter-bench) leads to a blog post + leaderboard comparing which LLM is the best at our robotic tasks.

30 comments

[ 4.3 ms ] story [ 58.5 ms ] thread
(comment deleted)
95% for humans. Who failed to get the butter?
I have a cat that will never fail to find the butter. Will it bring you the butter? Ha ha, of course not.
Someone actually paid for this?
The internal dialog breakdowns from Claude Sonnet 3.5 when the robot battery was dying are wild (pages 11-13): https://arxiv.org/pdf/2510.21860
I sort of love it; it feels like the equivalent of humans humming when stressed. "Just keep calm, write a song about lowering voltage in my quest to dock...Just keep calm..."
That is without doubt the funniest AI generated series of messages I have ever read.

Nearly as good as my resource booking API integration that claimed that Harry Potter, Gordon the Gecko and Hermione Granger were on site and using our meeting rooms.

ERROR: Task failed successfully

ERROR: Success failed errorfully

ERROR: Failure succeeded erroneously

ERROR: Error failed successfully

That was super fun - why is mine so boring ?
(comment deleted)
> The results confirm our findings from our previous paper Blueprint-Bench: LLMs lack spatial intelligence.

But I suppose that if you can train an llm to play chess, you can also train it to have spatial awareness.

will noone claim the Rick and Morty reference? I've seen that show like, once and somehow I know this?
>Our LLM-controlled office robot can't pass butter

was the script of Last Tango in Paris part of the training data? maybe it's just scared...

(comment deleted)
I guess I'm very confused as to why just throwing an LLM at a problem like this is interesting. I can see how the LLM is great at decomposing user requests into commands. I had great success with this on a personal assistant project I helped prototype. The LLM did a great job of understanding user intent and even extracting parameters regarding the requested task.

But it seems pretty obvious to me that after decomposition and parameterization, coordination of a complex task would much better be handled by a classical AI algorithm like a planner. After all, even humans don't put into words every individual action which makes up a complex task. We do this more while first learning a task but if we had to do it for everything, we'd go insane.

Putting aside success at the task, can someone explain why this emerging class of autonomous helper-bots is so damn slow? I remember google unveiled their experiments in this recently and even the sped-up demo reels were excruciating to sit through. We generally think of computers as able to think much faster than us, even if they are making wrong decisions quickly, so what's the source of latency in these sytems?
How can I get early access to this "Human" model on the benchmarks? /s
I wonder whether that LLM has actually lost its mind so to speak or was just attempting to emulate humans who lose their minds?

Or to put it another way, if the writings of humans who have lost their minds (and dialogue of characters who have lost their minds) were entirely missing from the LLM’s training set, would the LLM still output text like this?

It can't "lose" what it never had. :P A fictional character has a mind to the same extent that it has a gallbladder.

> if the writings of humans who have lost their minds (and dialogue of characters who have lost their minds) were entirely missing from the LLM’s training set, would the LLM still output text like this?

I think should distinguish between concepts like "repetitive outputs" or "lots of low-confidence predictions the lead to more low-confidence predictions" versus "text similar to what humans have written that correlates to those situations."

To answer the question: No. If an LLM was trained on only weather-forecasts or stock-market numbers, it obviously wouldn't contain text of despair.

However, it might still generate "crazed" numeric outputs. Not because a hidden mind is suffering from Kierkegaardian existential anguish, but because the predictive model is cycling through some kind of strange attactor [0] which is neither the intended behavior nor totally random.

So the text we see probably represents the kind of things humans write which fall into a similar band, relative to other human writings.

[0] https://en.wikipedia.org/wiki/Attractor

Funny I was looking at the chart like "what model is Human?"
The error messages were truly epic, got quite a chuckle.

But boy am I glad that this is just in the play stage.

If someone was in a self driving car that had 19% battery left and it started making comments like those, they would definitely not be amused.

The most surprising thing is that 5% of humans apparently failed this task! Where are they finding these test subjects?!
95% pass rate for humans

waiting for the huggingface Lora

Using an LLM for robot actuator control seems like pounding a screw. Wrong tool for the job.

Someday, and given the billions being thrown at the problem, not too far out, someone will figure out what the right tool is.

It feels misguided to me.

I think the real value of llms for robotics is in human language parsing.

Turning "pass the butter" to a list of tasks the rest of the system is trained to perform, locate an object, pick up an object, locate a target area, drop off the object.

when all you have is a hammer... everything looks like a nail