←
shao78 posts
shao avatar

shao

@shaomng

tron @uwaterloo, swe @magichourai (YC W24), eng @hackthe6ix

◷ Joined July 2020▧ 22 Photos and videos
shao avatar
shao@shaomng·Nov 9

Not towel-related, just a 4.0 elasticity/plasticity feature I found interesting: yield lets a sheet keep a crease. I pinned one edge of two identical stiff sheets, folded both in half the same way and let go. The elastic one sprang back open; the one with yield stayed folded.

0:07
shao avatar
shao@shaomng·Nov 9

Upgrading also drops the old 50-contact cap (it landed in 3.16), so cloth stops sinking into the table. I dropped a 1 kg ball onto the towel. On 3.14 it punched through to the table: the towel held 6% of its weight. On 4.0 it held 100%. With every particle in contact (486 contacts vs 50), the whole towel sits on the table instead of half of it sinking in, so the gripper has a real edge to slide under. That should make grasping easier.

0:03
shao avatar
shao@shaomng·Nov 9

4.0's sparse Newton solver ran my fold 3.1× faster: 16.8 s → 5.4 s for 5 s of simulation, with the same contact count. Newton takes a few big, curvature-aware steps instead of many small ones. Each cloth particle only touches its neighbours, so most of the system is zeros, and 4.0 skips them. on MuJoCo's 3-cloth drape demo it was 6.5×. For teleop, the towel's simulation went from 0.3× to 0.93× real time, so during teleop the towel now reacts in real time instead of in slow motion.

0:09
shao avatar
shao@shaomng·Nov 9

testing MuJoCo 4.0 on a towel-folding scene I built for @watonomous! this is a big advance for rollouts in deformable object manipulation. I scripted a trajectory that folds the towel in half, then ran the same scene on 3.14, 3.17 and 4.0 to see what changed.

0:05
Kuan Yi Wang avatar
⟳ shao retweeted
Kuan Yi Wang@kuan_yiii·Oct 27

calm project simulating humanity's next 10000 years in spatial conquest: - Seeded Markov sim, zoomable from the whole universe to one person. - 3-layer decoder-only transformer with handwritten backprop in TypeScript and no ML libraries. Tokens are decades. - PPO self-play league: the war AIs fight frozen earlier versions of themselves and get smarter over the centuries. -> apparently Dyson spheres aren't meta -> bars are still a thing in 11876 -> opus 100% grokked while training on the hunger games movies 💀 see if we reach Andromeda: 2soon2026.vercel.app/ plz don't roast it was for a hacky i believe in non-determinism built with goat @shaomng

0:52
shao avatar
shao@shaomng·Oct 19

ui: a live MuJoCo digital twin that mirrors the real arms' encoders

Post attachment
shao avatar
shao@shaomng·Oct 19

the hardware: two open-source SO-101 arms, an Arduino Mega, a $10 load cell, hobby servos and 3D-printed parts. ~$300 in parts

shao avatar
shao@shaomng·Oct 19

how it works: 📱 text a Photon agent 🧠 it plans the run and asks a human to approve ⚡ our bench controller drives 2 SO-101 arms plus an Arduino Mega running the carousel, sled and scale 🦾 arm paths planned in MuJoCo, fine-tuned by hand on the real bench

shao avatar
shao@shaomng·Oct 19

solution: a low-cost chem bench built from parts anyone can get. text it what you want, approve it, and the robots do the measuring, pouring and transfers.

shao avatar
shao@shaomng·Oct 19

problem: lab automation is expensive. the robots that measure, pour and mix for you cost a fortune to build and maintain, so most labs still have a person standing next to every reaction.

shao avatar
shao@shaomng·Sep 23

> hardware stack: - one camera per car, no encoders, no lidar, all run from one laptop

shao avatar
shao@shaomng·Sep 23

> our budget: - 3 robot cars at $50 each - went dumpster diving to build out the cardboard field

shao avatar
shao@shaomng·Sep 23

> features: - 30 Hz drive commands as fixed 32-byte UDP packets - Gaussian splats from 320x240 robot footage - one shared fleet map

shao avatar
shao@shaomng·Sep 23

> the solution: a harness that coordinates cheap robots in parallel as one fleet, sharing a single map of the space.

shao avatar
shao@shaomng·Sep 23

> the problem: deploying expensive robots is risky and costly, and current swarms lack coordination.