Applied Intuition is the physical AI company bringing intelligence to every moving machine on the planet.
Sunnyvale, CA
- Tweets 1,110
- Following 218
- Followers 6,550
- Likes 382
Pinned Tweet
Think big. That’s always been a driving motivator for us, and now we have a partner to match our ambition. Today, we’re proud to announce Applied Intuition’s strategic collaboration with HUMAIN to deploy physical AI across Saudi Arabia, starting with autonomous trucking. This is a major step toward our goal of making a billion machines intelligent.
Together, we will deploy thousands of autonomous trucks throughout key Saudi logistics corridors by 2030. This will make it the largest autonomous trucking network in the world, and our Self-Driving System and Vehicle OS technology will be the centerpiece of this collaboration with HUMAIN.
This is physical AI at national scale.
appliedintuition.com/blog/ap…
11
34
124
32,160
Applied Intuition retweeted
How can we exploit the large-scale in-the-wild dashcam data towards a driving foundation model? Our CVPR 2026 paper shows the power of such “free-gift” unposed and unlabeled data, learning driving representation with strong performance but significantly less labeled data!
Training AVs is complex and expensive to scale in new cities and terrains. From calibrated multi-sensor rigs to millions of labeled miles and painstaking annotation, we need to remove the blockers to autonomy development. Meanwhile, there are millions of hours of unlabeled dashcam footage.
And that’s where LFG comes in with a solution to exploit all this data.
🧵
3
8
1,518
Training AVs is complex and expensive to scale in new cities and terrains. From calibrated multi-sensor rigs to millions of labeled miles and painstaking annotation, we need to remove the blockers to autonomy development. Meanwhile, there are millions of hours of unlabeled dashcam footage.
And that’s where LFG comes in with a solution to exploit all this data.
🧵
1
2
14
2,415
The result:
→ On semantic segmentation, LFG surpasses its own teacher on future frames (0.751 vs 0.680 mIoU), despite predicting blind while the teacher saw the real images.
→ Depth predictions on future frames closely match the teacher (0.31 vs 0.26 AbsRel on KITTI-360, 0.22 vs 0.19 on Waymo Open Dataset).
→ It correctly separates moving objects from static background even when the automated labeling pipeline gets it wrong.
1
2
348
On NAVSIM, LFG with only 10% of labeled data (81.4 PDMS) already matches other pretrained encoders trained on the full dataset (PPGeo 74.6 PDMS, DINOv3 81.4 PDMS, Pi3 82.8 PDMS). With the full training set LFG reaches 85.2 PDMS, beating multi-camera and lidar systems like UniAD (83.4 PDMS) and Hydra-MDP (84.7 PDMS), using a single front camera.
arxiv.org/abs/2602.22091
167
When you're building a top tier team aimed at solving the most complex problems in physical AI, mediocrity is not an option.
Over the years, @malharhar has interviewed and hired many across Applied. As a high growth company, one of the biggest challenges has been maintaining and evolving the culture at scale.
review.firstround.com/firsth…
1
6
684
Our work with @dstlmod is shaping how the @BritishArmy can effectively deploy swarms of autonomous systems to deliver operational effect.
We led the Disrupter Group, bringing together Rowden, Evolve Dynamics and SAIF Autonomy to deliver this capability Test Bed in just a matter of months.
Learn more about our work: appliedintuition.com/defense…
Working with the @BritishArmy @Bluebearsystems and @AppliedInt we’ve successfully experimented with a collective control drone swarm in a realistic military environment
🔗 gov.uk/government/news/dstl-…
#Dstl #Autonomy #AI
1
1
5
1,222
Hundreds of drones, built by different vendors, tasked with one mission. Before they deploy, someone has to check the pieces work together, and today that usually happens at a live event, often the first time these systems actually talk to each other.
Applied Intuition's Digital Proving Grounds fix that — an always-on, vendor-neutral sim environment spanning the full sim-to-live continuum:
🔷 SIL: software tested continuously in a shared DevOps environment
🔷 HIL: real hardware verified against simulated environments
🔷 LVC: live platforms stress-test C2 at scale, no full fleet required
🔷 Live exercises: validate readiness, don't discover gaps
Applied Intuition partnered with the @DoWCTO on Virtual Readiness and Experimentation, integrating air, surface, and subsurface vehicles from different vendors on one mission, compressing the find-fix-fly loop from weeks to under an hour. Now scaling across the @CDAODoW's Autonomy Factory.
appliedintuition.com/blog/di…
2
540
Applied Intuition retweeted
Qasar Younis, CEO of @AppliedInt, has been a mentor and friend to @AshwinSreenivas since the earliest days of Decagon.
He came to our SF office to share what he’s learned across two startups, Google, YC, and 8+ years building Applied Intuition.
Thanks for joining us, @qasar!
3
3
39
10,729
We're headed to @UCBerkeley tomorrow, September 15th, for the Cal Engineering & Technology Career Fair.
Come by booth #21 to learn more about the physical AI problems we're solving at Applied, from autonomous vehicles and agriculture to trucking, defense, and beyond.
appliedintuition.com/careers…
3
9
1,217
Great code is only half the job. Making it work reliably on physical hardware in the real world is the other half, and it's the part most people don't get to touch.
The ability to drive impact and take ownership has been a big reason why our Deputy CTO, @malharhar, has stayed at Applied all these years. He has been able to go where the problems are.
Read his @firstround write up to learn more about his experience: review.firstround.com/firsth…
1
2
18
1,662
At this year’s Agentic AI Summit hosted by @BerkeleyRDI, our Chief Scientist, @Wei_ZHAN_ shared some of the cutting edge research he has been leading in the physical AI space.
As the industry races to deploy L2++ ADAS with imitation-learning-based E2E, Wei tackled a key question: what if end-to-end autonomy could be trained without imitation at all, relying only on reinforcement learning? An idea that runs against the current status quo.
Here are some of his insights
🧵
1
6
31
6,173
The takeaway: self-play / closed-loop RL here isn't post-training polish, it's a powerful pre-training tool, and framework throughput + world-model quality become the decisive factors for closed-loop scaling.
Autonomous driving is one of the earliest and largest at-scale production deployments of physical AI. If imitation-free, closed-loop RL can match and beat SOTA imitation methods here, many other physical AI domains can significantly benefit from closed-loop scaling. Open-loop scaling in AV 2.0 brought ADAS to the real world, and closed-loop scaling powered by large-scale RL and world models is driving robust autonomy that paves the way for AV 3.0.
1
165
Learn more about TerraZero here:
To obtain robust autonomy, it has always been the desire for planners to experience large-scale, high-value scenarios in ultrafast closed-loop training with minimal unit cost of data and compute, ideally self-evolving even without offline human demonstration.
TerraZero is our straightforward response at @AppliedInt to such a desire by scaling self-play reinforcement learning with procedural simulation and zero demonstration:
- Insanely fast in closed loop, i.e., 2.8 million simulation steps per second on 8 GPUs;
- Endlessly streaming hard and diversified corner cases with a procedural generation;
- Robust self-play recipe trading sample efficiency for compute efficiency;
- Zero demonstration in training, and zero-shot emergence across domains (cities, datasets, etc.);
- Minimal unit cost for experiencing high-value data points with a highly scalable framework.
So excited to see that one stack powered by TerraZero can serve various outcomes:
- Setting the state-of-the-art in the challenging, long-tail planning benchmark (InterPlan) with a clear edge;
- Topping the realism among demonstration-free methods in sim agent benchmark (WOSAC);
- Comprehensive practicality with policy covering heterogeneous agents (VRUs, trucks, trailers, etc.) and tackling noisy policy inputs.
On our website, you can also select arbitrary combinations of scenario/agent/view features and see how TerraZero works - play with it!
Really impressed by the potential of large-scale, high-value synthetic data in closed loop powered by ultrafast training, and TerraZero is just the beginning of our Terra Series research - committed to pushing the boundaries of physical AI learned in closed loop with scaled simulation and minimal demonstration. More follow-up work will be announced soon, stay tuned!
TerraZero report: cas-bridge.xethub.hf.co/xet-…
TerraZero website: terra-applied.github.io/Terr…
Terra Series website: terra-applied.github.io/
226