MOSS the 3D printed open source litter picking robot on the beach for the first time. Tracks are well sized for sand: dry sand low dunes, damp sand and the waterline, no problem. The big dry dunes near the entrance are another story, the ground clearance is low and... it got stuck twice. A quarter of the way down they get smaller and it passes without trouble. To take the big ones it would need a bigger robot or more ground clearance. It does not drive straight yet, I correct all the time, same as on the roadside video. That one is on me: the two tracks are not tensioned the same. Two tensioners per track, one sets the idler angle and the other the tension, and I only set them by eye after mounting. The encoders agree: on the road both motors turned the same within 1 %, so the drift is between the track and the ground, not in the motors. Battery: 12.1 V at switch-on, 11.8 V at rest back home. The 21700 cells are on their way. The rover was designed in Fusion with Codex in just a couple of weeks. OpenAI OpenAIDevsshow more

Build guide CAD & STL files Bill of materials GitHub repository Join the kit waitlist

You should visit @mehul and license his algorithms.

@mehul Matic (@maticrobots) is such a great bot! As an independent open source builder I don't think it's possible for me to licence something like this tho

@mehul @maticrobots You should meet anyway. And you never know until you ask.

@mehul @maticrobots Very true

I can imagine a ton of customers that want their grounds continually cleaned up. A night vision capability would timeshift the task to low traffic times.

My robot vac is good at cleaning with the light off, for litter pick and place I'd add a headlight tho. Interesting to explore this when daylight is nailed.

I just took a walk through San Francisco. Leaves falling on the sidewalks. Imagine if the robot were to pick them up, one by one, stack them like cards, and clear the block by morning?

Beach cleanup! Full size industrial robot arms mounted on autonomous ATVs with a dump trailer in tow is what I ultimately see here. You should see if you can get a sponsorship from one of the big guys and build @sendcutsend @OSHBuilt

We're gonna need a bigger version!

Agreed, but you know how scaling a robot also scale it's cost: Back-of-the-envelope calculation 2 x 2 x 2 = 8 4 x 4 x 4 = 64 8 x 8 x 8 = 512 If a 20 cm robot cost $80 A 40 cm one cost $640 A 80 cm one cost $5120 Ofc it's not linear like this as some parts remain the same price. But it's the idea.

This is so cool. I want to see it pick up some gahbahge!

Me too 😆

Following to see!

WOAH i love this guy! would love to build and send to some NYC beaches

Sand is a serious terrain challenge! Well done. Did you have to take it apart to clean it? Next, the chasm between remote control and autonomous..! You might encounter me down in the gorge.

Awesome this is the idea of opening it, there’s only so much I can do. Nope just blasted some compressed air.

Compressed air cleaned = well designed. Bet mine would be crunchy after the beach. My next move would be to choose a specific use case and setup a training loop. First figure out how it decides where to go next. Need to trust it won't go where it shouldn't. Then can set it on autopilot and dive into trash identification: probably fine tune a SAM. Then collect teleop data on how to pick up different types of trash, fine tune whatever the latest/greatest action/world model is. Get to a supervisor level with the right buttons to catch edge cases and make corrections that can be fed back into training. Then see how many you can supervise at once, how hands off you can be. Finally leave that system on auto and jump to the next use case. I looked at the unknowns on that timeline and decided to do a fundraising project: augmented reality multiplayer games played on your phone's web browser.

That's a big road map. I see you have your own rover, I thought you was building MOSS, awesome let's share our notes, I am just getting started, I am trying to avoid overthinking it, the part that will need the most training is the detector I think, then aim the gripper at the trash and the pick and place should be a kind of scripted reflex at some point. I don't like much how stuff is complexified with AI lately. I don't think a litter picking robot need this. If the grasp fail try again, it's cheaper than LLM orchestration of VLA via API kind of stuff.

Yeah, I landed on that same differential drive tank-tread design in the spring. The difference between what I made playing with blender and what you made is humbling! Your components all fit and stuff.. Overthinking it comes with trying to get automation to work. "If the grasp fails try again" is an ACT policy. I had about 70% success picking up a specific object in a specific environment with just 30 training examples. Can run the training on an old gaming PC and run the policy itself on a Raspberry Pi. No APIs. Your sensor stack rocks. DimOS is the future. Running DORA on the bot's Pi-5 instead of ROS makes life better. Can fine tune a vision model that should run on the realsense (runs on my Oak-D Lite with depth). That's a free-software-on-an old-gaming-computer job too. Lots of labeling. So it's: - trash search - identify trash - deterministic move in reach - fire the ACT policy depending on the trash - deterministic drop in bucket - rerun loop. I didn't know anything about this a year ago. You got this.

The button supervision make me think of DAgger, I've got 90% on ACT a few month ago but from my perspective there is too much variability in litter and conditions for ACT / VLAs, so I start with a scripted grasp and the wrist camera as detector. Do you know a good labeled dataset for litter ? If I don't find one I may spend a couple day shooting litter form different angles and have a model like GroundAnything, from GroundingPI to label the data.

Right, your on it! Excited to see if a scripted approach works, cuz yeah, variabilities.

Mee too lamo, I'll report on X for sure, mabe I am wrong but let's try the less expensive first.

Would love to manage a fleet of these! We need a Trash-Collected-Per-Moss leader board.

Equal motor rotation doesn't guarantee equal travel over sand. Comparing track deflection under the same applied force could give you a repeatable tensioning baseline before adjusting the steering controller.

This is a very complicated way of saying I need to adjust tracks tension.

Littering the landscape with litter picking robots. 100 years from now even the bugs will be robotic.

how much sand did you have to clear out afterwards? a quick post-beach teardown would be fun to see.

Nothing inside the hull, I just use the air compressor on and inside the tracks.

How do you deal with waterproofing/Splash proofing? Or does your design not accommodate for that

Yes I have tried to design it splash proof, lips at the hull / cover interface, non woven membrane for the speaker, 45° vents for the fans, I've printed a set of TPU gaskets but did not installed it yet. I love all these questions about collecting litter at night, under the rain, but thb from my perspective, I'll do it in good conditions first, there is still a lot a work on the autonomous side. It's the current priority. For now it should be good enough to handle splashes, not rain yet, arm servos would not like this too.

Agreed with you. My robot is using a similar stack to yours for a different function. I want to make it workcij ideal conditions first and work on autonomy then the weather proofing can happen around that