Emerson Segura

CTO,ML,Research

[Emerson Segura @emerson]

@kscottz agreed, a lot of work left to be done

[Emerson Segura @emerson]

@jana_iris Same, I got it back.. the 4th time now.. in Tenderloin for $30 - I had an apple tag on it.. this is the 5th time my bike was stolen...and 4th recovery (so lost 1 bike in about 3y)

[Emerson Segura @emerson]

@TongheZhang01 unlikey in 12m- yet when that does happen, mills. of people won't be needed to assemble tiny iphone parts (glue, screws,fragile zif connectors) and robots will be able to free surgens from the delicate and squishy physical aspect of surgery - I look forward to that future

[Emerson Segura @emerson]

@thehumhuworks @TongheZhang01 Not really: try using computer vision to do surgery... or figure what hydraulic line does what in mechanism such as landing gear to say petroleum refinery.... humans in those fields can figure it out.. no easy/solved way to that with CV/LLms yet

[Emerson Segura @emerson]

@keerthanpg Agreed for practial applications for the next ~10y: more data, more work that can be automated.

(yet some people also extrapolate this to AGI/human level inteligence/conciousness & it does not seem any ML is anywhere near the refined machinery that is life,and animal cognition.

[Emerson Segura @emerson]

@ZeeshanZiaML very insightful observation

[Emerson Segura @emerson]

@earthtojake Lol

[Emerson Segura @emerson]

@gaurangm15 @katieruthmishra agreed,also the trust and back&forth with the customer is criticall for sucess, women don't just trust anyone with their hair and may freak out if there are issues. People doing this have the social skills to manage this-like a high end resto: it's not just about perfect cooking

[Emerson Segura @emerson]

@ErenChenAI probably got a deal - heard realestate prices there have been going down every year , since 2020.

[Emerson Segura @emerson]

@TheHumanoidHub is the fruit company still working on a self driving car or is it a siri humanoid now their main robotics project? :)

[Emerson Segura @emerson]

@TiffanyFong it was the leather jacket, she was really into that.. nothing to see here..

[Emerson Segura @emerson]

@TheHumanoidHub could you try something fun, like "place the black box in the grapes", or "place all the objects as far away from table as possible" :)

[Emerson Segura @emerson]

@DrJimFan the $50k hands look nice, looking fwd to groot2

[Emerson Segura @emerson]

@AndrewCurran_ nice, yet $50k hands -- looking fwd to Groot2

[Boris Belousov @_bbelousov]

MuJoCo 4.0 from Google DeepMind is the first major version of the physics engine since 3.0, three years ago. Its headline feature, speculative contacts, keeps thin and fast objects from passing through each other, which allows much larger timesteps.

The problem

• Collision detection checks geom poses only at discrete timesteps. A thin or fast object can move further than its own thickness in one step and tunnel through, so the usual fixes are a smaller timestep or an inflated margin.

How it works

• Speculative contacts: each fast-moving geom is swept forward through the step to find when it would hit something. The solver gets a contact like any other that stops the motion at the surface, with no sub-stepping.

• Deformables are now solved implicitly together with rigid bodies, joints and tendons, which keeps cables, cloth and shells stable at 5-10x larger timesteps. New are penetration-free cloth and elastic rods with bending and twist.

• Plasticity: wires, clay, folded paper and metal foil now keep their dents and creases.

Does it pay off?

• In the team's tests, an arm catches a 0.7 mm playing card falling at over 2 m/s with a 16 ms timestep every time, where discrete collision detection misses it above 8 ms.

• Humanoid stair climbing and thin-object grasping run 6 to 9 times faster, because they can use larger timesteps.

Also in 4.0

• MuJoCo Studio, a new viewer SDK with physically based rendering. The same tools run in a browser tab with nothing to install.

• MuJoCo Playground 1.0 ships alongside.

Caveats

• The speedups and timestep factors are the team's own measurements, without benchmark details in the announcement.

• Speculative contacts predict motion from pre-step velocities, so a slow geom pushed across an obstacle within one step can still tunnel. They cover convex primitives and meshes on CPU. Heightfields, SDFs, flexes, MJX and MuJoCo Warp are listed as coming next.

• Breaking changes: composite bodies, the cable and shell plugins and flexvert equalities are removed, Python 3.10 is dropped, and 3.x installs need a pip uninstall first.

By @yuvaltassa and the Google DeepMind Robotics Simulation Team.

[Emerson Segura @emerson]

@withplayground LOL! so your Astra ran without internet- great then pls share the Astra weights then so we too can run it "without da internetz"..

[Emerson Segura @emerson]

Skynet is going to remeber this !

""Data centers rejected by voters across the country —
challenging the AI boom

The nationwide data center buildout has been at the heart of the U.S. economic growth story , driven large stock gains ""
https://t.co/VJPaoO1699

[Emerson Segura @emerson]

@SihangLi659492 Nice work

[Emerson Segura @emerson]

@CJHandmer Presumably all those rfics have well documented spec sheets and reference designs, thats the norm. Still cool, and very useful,

(Yet 95% of the work was done inside the chipsets, with experienced people optimizing them over decades, and in fabs costing billions of dollars).

[Emerson Segura @emerson]

@SCH_Clay good,what does 15 followers have to do with it... tonnes of competent people are not on twitter