Akash Sharma (@akashshrm02) — posts, photos & videos

@akashshrm02 reposted

@keenanisalive

With everyone busily using LLMs + Blender/Three.js to make 3D models, I thought I'd try a different geometric representation.

This dragon isn't a mesh, NeRF, 3DGS, or video model: it's 27kb of GLSL code (generated via LLM), defining a closed-form implicit surface. 🧵 [1/n]

0:12

87 275 36 3.7K 313K 3.2K

Akash Sharma

@akashshrm02

Oct 8

Object trajectory tracking is pretty sweet as an interface for policies to learn!

Checkout @yulinnliu07 ‘s awesome work!

Yulin Liu@yulinnliu07 · Oct 8

🤖 For parallel-jaw grippers, high-level plans can already be executed with mature primitives such as grasping. What about dexterous hands?

Introducing GOTT (General Object Trajectory Tracking): object trajectories to dexterous execution with a reusable cross-embodiment contact primitive.

The same primitive supports diverse real-world tasks including sweeping, pouring, whiteboard erasing, and hammering.

With one checkpoint, GOTT reaches 90-96% grasp-and-lift success across multiple objects on three real-world arm-hand platforms.

Check videos 👇

0:31

Open quoted post →

0 1 0 11 737 1

@akashshrm02 reposted

Yann LeCun

@ylecun

Oct 7

That's a commonly ignored fact.

It's not like Leonardo da Vinci invented the helicopter, or even Paul Cornu who built one that took off in 1907.

But Louis Bréguet invented the swashplate, and Henrich Focke flew a controllable helicopter with contra-rotating rotors, all before Igor Sikorsky came up with the variable-pitch tail rotor and made the whole thing practical in 1939.

Getting the basic idea is the easy part.
Building it is harder.
Making it work is even harder.

24 45 9 511 60.2K 96

@akashshrm02 reposted

Unnat Jain

@unnatjain2010

Oct 7

Objects in robot demos are awfully patient 🙃

We’ve spent months on tasks where they aren’t. Catching balls, pouring into a moving pot, taking a bottle from someone walking by.

Two additions helped π0.5, RTC, & VLASH handle dynamic manipulation! More from our CoRL'26 oral in 🧵

Daniel & Jimin ran 1,500+ physical robot evaluations to put it to the test, with 50 runs per reported success rate.

As always, we’re making everything public: tempo-robot.github.io

0:48

16 35 4 181 9.7K 98

@akashshrm02 reposted

Kenny Shaw

@kenny__shaw

Oct 5

Don't forget to submit to our 3rd workshop on Dexterous Manipulation at CoRL 2026 in Austin! Submissions close tonight!

dex-manipulation.github.io/corl2026/

Co-organized with @HaozhiQ @Ying_yyyyyyyy @akashshrm02 @SharpaRobotics @deepakpathak @JitendraMalikCV

0 3 0 24 1.1K 1

Akash Sharma

@akashshrm02

Oct 1

Perhaps counterintuitive, but this is fundamentally true.

IK is built on a simple model of the arm and doesn’t account for things like friction, arm dynamics, environment. With RL you can get reactivity, distill to choose the right IK branches to avoid singularities.

Bitter lesson much?

Qiayuan Liao@qiayuanliao · Oct 1

Any online optimization (IK, ID, MPC) for low-level control should be abandoned. A RL policy can fundamentally track more precisely, more robustly, and adaptively to unknown disturbances and uncharacterized noise, especially for low-cost hardware, even on a manipulator.

Open quoted post →

0 2 0 32 3.5K 9

Akash Sharma

@akashshrm02

Sep 27

I’ll be giving a talk at the sensors and actuators IROS workshop at 2:20pm today.

I will talk about what I think are, “the missing ingredients in scaling dexterous manipulation data”.

Workshop link: sensact4dexman-workshop.github.io

Come by and say hi if you’re in person!

2 3 0 27 1.3K 1

Akash Sharma

@akashshrm02

Sep 26

Love it! This looks awesome for someone that wants to stay on Linux, and also fixes many pain points with existing slide editing software.

Vincent Sitzmann@vincesitzmann · Sep 24

I am open-sourcing DeckWerk, a multi-platform slide editor with stellar video support, real-time collaboration, and Agent support. I used to be on Linux but had to switch to Mac b/c there was no good enough slide editor - @OmarchyLinux & @dhh inspired me to vibe-code my own, so I'm finally back on Linux!

Install from source via GitHub (installers coming soon!):
github.com/vsitzmann/deck…

11:08

Open quoted post →

0 0 0 4 1.7K 1

Akash Sharma

@akashshrm02

Sep 20

Things are starting to actually, actually work! Such exciting times :)

Nick Maselli@NickBuilds11 · Sep 18

Robotics is truly the most exciting industry in the world right now

Open quoted post →

2 0 1 24 3.3K 2

Akash Sharma

@akashshrm02

Sep 13

If you're working on human to robot learning, we'd love to see your work!

0 0 0 2 243 0

Akash Sharma

@akashshrm02

Sep 13

Dexterity, tactile intelligence and perhaps distillation into System 1 reactivity is the next frontier.

Aviv Tamar@AvivTamar1 · Sep 13

Just a reminder - robotic touch is not solved, and there is not much data to train on (currently). So there's still interesting work to do.

Open quoted post →

1 1 0 29 3.2K 12

Akash Sharma

@akashshrm02

Sep 7

PTLD is accepted at CoRL 2026!

Distilling from privileged sensor policies in the real world has been proving to be a sensible strategy times over in follow up projects.

Code release coming soon!
We will also release an updated multiview foundation pose tracker that is more generally useful.

Akash Sharma@akashshrm02 · Jul 11

This was my final Ph.D. work co-led by Rosy Chen and I.

Prior work fixates on zero-shot sim-to-real resulting in compromises such as the choice of simple sensors like proprioception.
With PTLD we show that if you're allowed to collect a little real data, you can actually deploy rich sensorimotor policies with significantly stronger performance.

Paper: PTLD: Sim-to-Real Privileged Tactile Latent Distillation for Dexterous Manipulation.
@rosychen0501, @mukadammh, @michaelkaess, Tingfan Wu, Francois Hogan,@JitendraMalikCV, @akashshrm02

CMU · UW · UC Berkeley · FAIR at Meta.

📄 arxiv.org/abs/2603.04531
🌐 akashsharma02.github.io/ptld-website
🎥 youtu.be/YVFZt2YAS3Y

0:16

Open quoted post →

9 2 0 76 7K 20

Akash Sharma

@akashshrm02

Jul 13

Checkout @Milotrince 's work on implementing tactile sensors in Genesis simulator.

Core idea is to use SDFs / raycasting from predefined surface probes to measure contact + forces + torques, along the lines of hydroelastic contact models.

Trinity Chung@Milotrince · Jul 13

8/ For a more complete discussion on my work in tactile sensing, my master's thesis presentation is available at youtube.com/watch?v=YP7WUN… 😄

Open quoted post →

0 0 0 11 1.7K 4

Akash Sharma

@akashshrm02

Jul 13

I'm at RSS this year in Sydney!

I've been thinking a lot these days about tactile sim-to-real, exoskeletons, forceful manipulation, and tactile retargeting to leverage human data. Dms open and always open to talk about these topics. I believe the next frontier of robot learning is forceful.

We're also presenting TactAlign (lead author @WiYoungsun) today in manipulation session 1 (July 13, 2:30pm - 5:15pm).

2 3 1 21 6.6K 3

Akash Sharma

@akashshrm02

Jul 11

Presenting PTLD: an approach to learning tactile dexterous policies without ever simulating the tactile sensor.

Tactile is essential for performing highly dexterous manipulation. However, collecting tactile observations reliably has been a fundamental bottleneck: (1) teleoperating a multi fingered hand for dynamic tasks is challenging, making sim-to-real imperative, (2) one can’t realistically simulate tactile today

0:10

6 46 10 191 38.4K 109

Akash Sharma

@akashshrm02

Feb 26

Late life update: I defended my Ph.D. @CMU_Robotics last month, advised by @michaelkaess.

This culminates over 4 years of chasing dexterity with robot hands. I'm grateful to my supportive committee: @shubhtuls , @GuanyaShi , @mukadammh & @JitendraMalikCV

3 1 1 101 4.9K 10

Akash Sharma

@akashshrm02

Feb 17

It's easier to collect human demos over robot demos, but how can we make the best use of this data?
Check out our work: TactAlign!

In TactAlign, we not only retarget motions, but also tactile readings via rectified flow. Many policies here are trained purely on human data!

Youngsun Wi ✈️ RSS 2026@WiYoungsun · Feb 17

Dexterous hands vary widely—so do tactile modalities. 🖐️🌈
Our vision on tactile human-to-robot transfer:
🔓 Not tied to specific hardware
♻️ Reuse human tactile demos across embodiments

Presenting TactAlign, a cross-sensor tactile alignment for cross-embodiment policy transfer.

0:11

Open quoted post →

0 5 0 16 2.9K 8

Akash Sharma

@akashshrm02

Dec 12, 2025

Super excited to finally release Any4D!
It's been more than a year in the making with a lot of hard work from @JayKarhade
Any4D is feed-forward so it is orders of magnitude faster. It's not query based, so it can handle dense scene flow.
It will only get better with more data!

Jay Karhade@JayKarhade · Dec 12, 2025

Introducing Any4D, a unified transformer for fully feed-forward, dense, metric-scale 4D reconstruction from flexible inputs!

Any4D regresses per-pixel motion + geometry across frames in one pass — 15× faster, 2–3× more accurate reconstructions ⚡📈

Details + code below 👇

Exciting collab with @Nik__V__ @YuchenZhan54250 Tanisha Gupta @akashshrm02 @smash0190 @RamananDeva

0:44

Open quoted post →

2 6 0 39 3.5K 11

Akash Sharma

@akashshrm02

Dec 1, 2025

Didn't expect to be out so soon @carlo_sferrazza ! Awesome stuff, looks like an excellent starting point for humanoid research.

Carlo Sferrazza@carlo_sferrazza · Dec 1, 2025

Sim-to-real learning for humanoid robots is a full-stack problem. Today, Amazon FAR is releasing a full-stack solution: Holosoma.

To accelerate research, we are open-sourcing a complete codebase covering multiple simulation backends, training, retargeting, and real-world inference.

0:34

Open quoted post →

1 0 0 6 727 0

Akash Sharma

@akashshrm02

Nov 21, 2025

How can one make a robot do things from just observing humans?

- Use the awesome Aria 2 glasses + FoundationStereo for depth & extrinsics.
- Point policy to get 3D generalization for free.
- Just 1 onsite human demo

This was a fun project led by our dedicated @irmakkguzey!

Irmak Guzey@irmakkguzey · Nov 21, 2025

Dexterous manipulation by directly observing humans - a dream in AI for decades - is hard due to visual and embodiment gaps.
With simple yet powerful hardware - Aria 2 glasses 👓 - and our new work AINA 🪞, we are now one significant step closer to achieving this dream.

0:59

Open quoted post →

1 4 0 10 1.1K 1