Building in physical AI | Ex-NVIDIA, Ex-Microsoft, PhD AI/ML @ NYU
New York, NY
Joined November 2010
- Tweets318
- Following1.4K
- Followers222
- Likes1.9K
I’ll be at #CoRL2026 this week, if anyone wants to talk common sense manipulation or adapting policies to unknown environment or where to find good coffee here or anything else hmu!!
2
144
Adversarial attacks are supposed to be imperceptible to human eye. If you change the input image, the output would change as well. 🤷♂️🤷♂️
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
111
Congrats to @moonlake ! Purpose build models and agents still have an edge over general models like astra for several tasks but the gap is definately getting smaller. 🫣👀
Been seeing a lot of Astra for real2sim, but what you need for robotics: physics and affordance.
Our SOTA recipe specifically solves for:
• Affordance: model object details that convey functionality.
• Physics: calibrate simulation dynamics against real-world observations.
Our agent has the action space of a “full-stack simulation engineer” and jointly considers geometry, materials, colliders, articulation, and physics. It automatically calibrates physics against real trajectories and matches task-specific dynamics.
Simulation environments is useful only if it can provide meaningful policy level signals. Methods, comparisons, and results below.
3
136
Instead of “encode/condition on the prompt” we need to do “tool call, web search, retrieve and then condition” for diffusion models. @OpenAI #DiffusionModels #GenAI
79
Devansh Bisla ✈️ (CoRL - 2026) retweeted
We're releasing LAION-BVD: a 10-million-hour open video dataset for multimodal pre-training.
- 1.3B video URLs from CommonCrawl
- 80M downloaded videos
- 10M video hours
- 55M captioned clips
- 300M frame-caption pairs
🌐: projects.laion.ai/bvd/
Devansh Bisla ✈️ (CoRL - 2026) retweeted
New work with @nvidia: evaluating robot policies entirely inside a world model. The policy acts, the model imagines the consequences, and the imagined evals predict real-world results. 🧵
real vs world-model rollout side by side📷
Devansh Bisla ✈️ (CoRL - 2026) retweeted
Gemini Embedding 2: A Native Multimodal Embedding Model from Gemini 🚀
Today, we’re sharing the @GoogleDeepMind white paper for GE 2, our first native multimodal embedding model. Whether it’s text, audio, video, or image, GE 2 provides a unified representation of the input.
Devansh Bisla ✈️ (CoRL - 2026) retweeted
I trained an autoencoder that reconstructs images with zero reconstruction loss.
No MSE. No image space supervision.
The only signal: "According to you, does your output look like your input through your own eyes?"
It works.
Blog link, demo and summary 👇
Devansh Bisla ✈️ (CoRL - 2026) retweeted
Demo night, @spc NYC, May 20
Lineup
@LumaLabsAI — Ray3 (reasoning video model)
@ElevenLabs — ElevenMusic (studio-grade AI music)
@SentienceCom — personal AI “digital twins”
@sandbar — voice ring for notes/actions
@floraai — unified creative AI
Apply below
Probably the lowest piece of argument clinic in this week of @TheEconomist.
The piece poorly tries to wash away the history of Mughal rule over India in guard of giving us paneer and biryani? What? 🤷
74
Devansh Bisla ✈️ (CoRL - 2026) retweeted
Today, we released Lyra 2.0, a framework for generating persistent, explorable 3D worlds at scale, from NVIDIA Research.
Generating large-scale, complex environments is difficult for AI models. Current models often “forget” what spaces look like and lose track of movement over time, causing objects to shift, blur, or appear inconsistent. This prevents them from creating the reliable 3D environments required for downstream simulations. Lyra 2.0 solves these issues by:
✅ Maintaining per-frame 3D geometry to retrieve past frames and establish spatial correspondences
✅ Using self-augmented training to correct its own temporal drifting.
Lyra 2.0 turns an image into a 3D world you can walk through, look back, and drop a robot into for real-time rendering, simulation, and immersive applications.
➡️ Learn more: research.nvidia.com/labs/sil…
📄 Read the paper: arxiv.org/abs/2604.13036
Protecting interest of one group doesn’t automatically mean harming other.
He also didn’t say I will not let robotaxis operate period. Although, it will be a terrible decision.
The question is asked at 15:00
xitter.gq/nycmayor/status/203904…
Congrats to NYC Mayor on yet another push to harm the majority in the interest of protecting a minority interest group
Waymo rollout in NYC has been stopped
85
My feed is talking about custom CUDA kernels, so I started digging in last week. I’ve put together a "fill in the missing code" guide for reproducing GPT-2 here: github.com/devansh20la/repro…
The boilerplate & paths are wired up-you just implement the forward paths with kernels.
1
88
My first pass is on the main branch (```git checkout main```). There’s still plenty of memory and synchronization overhead to iron out for anyone looking to get closer to the hardware.
1
50
Damnnit! I just started learning it but seems like we are going back to BASIC now:🤣
xitter.gq/blelbach/status/203937…
Today, NVIDIA is launching the next paradigm shift in GPU programming: cuTile BASIC
Write perf portable BASIC kernels and deploy them at any scale from edge inference devices like your calculator to entire GPU clusters
We're going back to BASIC
developer.nvidia.com/blog/cu…
46