Wessam

Founder @PolymorphInt

[Elon Musk @elonmusk]

SpaceX option package for new Tesla Roadster will include ~10 small rocket thrusters arranged seamlessly around car. These rocket engines dramatically improve acceleration, top speed, braking & cornering. Maybe they will even allow a Tesla to fly …

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[Wessam @wessam_hamid]

A machine shop is full of machines that could do much more if they didn’t need humans to specify every variation. Give them a model that understands the part, fixture, and task, and the same hardware can handle far more.

[Wessam @wessam_hamid]

If a model ports to an excavator with an hour of data, the moat moves from the model to whoever has access to the machines.

[Sergey Levine @svlevine]

π-0.7 API hackathon
where we learn that in fact an excavator and a nerf gun are also robots

kudos to @ActorLabs for the awesome event

[Wessam @wessam_hamid]

Once the brain is portable, every machine with an actuator is a candidate. That's a much bigger market than "robots."

[Wessam @wessam_hamid]

@Alpha10six @OpenAIDevs Are the lights sharing a CAN bus with the joints or running on their own?

[Wessam @wessam_hamid]

Action chunking works better than anyone can fully explain. Predict 50 steps, execute a few, re-plan. Half of modern imitation learning depends on it.

[Wessam @wessam_hamid]

An hour of continuous autonomy is a different milestone than 95% on a task. Long runs surface what short evals never see: drift, heat, a cable slowly working loose.

[Kun Huang @kun_h____]

Here’s a more personal breakdown of how we built Dyna-2.1 - the first Physical Agent to achieve 1-hour-long, dexterous whole-body autonomy.

A lot of the system came down to a few fairly fundamental design choices.

[Wessam @wessam_hamid]

@DataScienceHarp Which of the six checks flags the most episodes on real teleop data?

[Wessam @wessam_hamid]

Give the same unified model wheels, legs, wings, or a track. I'm increasingly convinced the interesting part is what comes after that.

[Wessam @wessam_hamid]

A CNC machine can make almost anything. Forklifts can move almost anything. The bodies for much of the world's automation shipped decades ago. They just need a brain to decide what to do next.

[Wessam @wessam_hamid]

@SongShuran Congrats @JuntaoRen and @SongShuran on NEEDLEWORK!

[Wessam @wessam_hamid]

@cixliv You guys just keep raising the bar, and it’s amazing.

[Wessam @wessam_hamid]

Scripted collection teaches the robot your task list. Freeform teaches it what hands actually do between tasks. Not surprised which one scales.

[Jannik Schilling @Jannikschg]

In 8 weeks, we scaled our UMI data collection to 1M+ diverse tasks at $10 per hour of data.

How? Scripted tasks have low throughput, so we made a bet: collect massive amounts of freeform data, with operators interacting with objects on a table however they like. That gave us 5x+ more data than scripted collection.

The catch: freeform data is unusable without annotation, and VLMs couldn't annotate at scale yet when we started Pantheon.

Last month, VLMs finally got good enough. The bet paid off.

Check out the full blog post for more on our hardware, infra, and how a core team of 5 scaled to 90 operators in a few weeks!

[Wessam @wessam_hamid]

@radbackwards Building a better world is the only reason worth getting up early for!

[Wessam @wessam_hamid]

@chris_j_paxton Congrats @pruthvigeedh_ on the open gripper!

[Wessam @wessam_hamid]

@adcock_brett Huge first day Brett and congrats to the whole team!

[Wessam @wessam_hamid]

@DrJimFan The giveaway would be the cups. A robot puts every one back in the cabinet, a human leaves three on the counter.

[Wessam @wessam_hamid]

@cetiennec Where are you seeing it most, university labs or home builders?

[Wessam @wessam_hamid]

The most useful sensor on a robot is the one that tells you the other sensors are lying.