World Archive (@worldarchiveai) header
World Archive (@worldarchiveai) profile picture

Helping robotics labs reach AGI faster!

India
Joined June 2026
Photo by World Archive (@worldarchiveai)Photo by World Archive (@worldarchiveai)
World Archive retweeted
One thing from Dario Amodei’s latest essay that really stood out to me for robotics is the idea that capability and safety can’t move at completely different speeds. As robots become more autonomous and start operating in the physical world, we’ll need much more time spent on testing, evaluation, monitoring and understanding how they behave in edge cases. He makes a great analogy to other safety-critical systems: “There is precedent for operating technologically complex, safety-critical systems millions of times without anything going wrong.” The hard part isn’t getting a robot to work once. It’s getting it to work safely, reliably and predictably millions of times.
1
4
119
Robots learned to see the world. Now they’re learning to feel it. 30K+ hours of tactile data, 1.4M episodes, 450+ tasks. Vision + touch = embodied intelligence.
Do robots really need touch, or are we overcomplicating the problem? N⁰-Foundation makes the case for tactile intelligence with 30K+ hours, 1.4M episodes, 6 embodiments and 450+ tasks. The thesis is simple: vision tells a robot what is happening. Touch tells it what’s happening at the point of contact — force, friction, slip, deformation. But maybe better vision + better models can learn all of this without tactile sensors. That’s what makes N⁰-Foundation interesting: it turns “robots need touch” from a belief into something we can actually benchmark.
2
110
World Archive retweeted
Deployment is the eval. Every robot in the wild becomes a continuous source of data you simply can't reproduce in a lab. But deployment comes with a huge operational cost — and that's the hard part: Fleet management, maintenance, calibration, remote intervention, model updates, monitoring, uptime, safety — and eventually insurance when robots perform high-value or mission-critical work. If you can build the infrastructure to manage this, something incredibly powerful happens: - Every deployment makes the next model better. - Every failure becomes an eval. - Every eval becomes training data. - Every new deployment expands the distribution. This creates a massive new infrastructure vertical for robotics. Companies that can manage this entire loop for robotics companies could become the picks-and-shovels winners of the robotics era. cc @DynaRobotics
2
3
12
555
World Archive retweeted
The interesting part here isn’t the 2.4× speedup. It’s the change in how we think about VLA inference. Instead of treating each action chunk as an isolated prediction, FlashVLA turns decoding into a continuously running process. Feels like an important step toward making VLAs behave more like controllers than generative models.
VLA inference has a weird bottleneck: the robot spends most of its time waiting for actions to finish decoding. FlashVLA turns that into a stream. Instead of fully denoising one action chunk before starting the next, it keeps multiple chunks at different noise levels and refines them continuously. The result: action decoding becomes something the robot can keep running, not something it has to keep waiting for.
2
2
26
1,434
The real value of Microduck isn't the robot. It's the ability to cheaply generate physical failure data.
$399 for an open-source biped with the full RL + sim2real stack is kind of insane. The interesting part isn't collecting walking videos. It's measuring the adaptation cost: how many real-world failures/trajectories does a policy need to go from sim → real → robust? Could be a cheap physical AI benchmark hiding in plain sight.
1
113
World Archive retweeted
We’re spending a lot of tokens on robotics harnesses right now — and starting to simulate what this layer could actually become. Harbor using harness to automates robot RL across 6 benchmarks / 16 tasks spanning manipulation, locomotion & bimanual dexterity. Zetta reports 90.8% on LIBERO-Pro and 93.6% on RoboCasa with a closed-loop embodied harness. But these feel like the first primitives. Imagine a harness sitting above manipulation, locomotion, navigation, grasping, bimanual, dexterous & whole-body policies — managing context, deciding what runs when, verifying outcomes, recovering from failures and learning from every rollout. Then take it beyond one robot: coordinate a fleet, share state, reallocate tasks and recover when one robot fails. The VLA gives the robot a skill. The harness could become the system that turns skills into a coherent system.
The next big robotics layer emerging is the harness. NVIDIA is building agent harnesses. Zetta is building a closed-loop embodied harness, while Ludo uses a harness to connect VLM reasoning with memory, speech, navigation & manipulation. A new layer is forming between foundation models and robots: managing context, tools, execution, verification, recovery and feedback. The model is becoming the brain. The harness is becoming the operating system for physical intelligence. And the real moat could be the loop: every robot action creates data that makes the system better.
1
2
13
1,967
World Archive retweeted
The next big robotics layer emerging is the harness. NVIDIA is building agent harnesses. Zetta is building a closed-loop embodied harness, while Ludo uses a harness to connect VLM reasoning with memory, speech, navigation & manipulation. A new layer is forming between foundation models and robots: managing context, tools, execution, verification, recovery and feedback. The model is becoming the brain. The harness is becoming the operating system for physical intelligence. And the real moat could be the loop: every robot action creates data that makes the system better.
6
14
3
119
11,653
World Archive retweeted
great things happening in robotics sphere
Introducing GEN-1.5, a one-shot learner. It can learn new tasks in a few seconds. Show it what to do, and it generalizes. This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.
11
17
6
530
44,570
We have already indexed 100k hrs of HuggingFace data to detect duplicate. Aim is to reach 500k hrs very soon
35% of OTS deals in egocentric data might involve fraud. Vendors send a perfect sample. The final dataset is either full of recordings you already own or recordings already open-sourced on Hugging Face. Sellers can Rename, Trim, Crop, Re-encode. Checksums change. The content doesn’t. We just open-sourced Overlap. It indexes any OTS data and compares it against your inventory & public HuggingFace datasets. Ask sellers to share indexes first. Detect duplicates without ever seeing the real content. No GPUs needed. A normal MacBook can index 5k hrs data in an hour. Let’s change the industry: share index files instead of just 5mins samples. Stop paying for free or repeat data. Github - github.com/World-Archive/ove… GUI - overlap.worldarchive.co
2
227
World Archive retweeted
Robotics data needs a plagiarism checker. Not for the model. For the dataset. A vendor can send you a perfect 5-minute sample, while a large part of the dataset you’re buying may already exist in your own inventory or in public datasets. Before buying more data, we need to know how much of it is actually new. This is a step in that direction.
35% of OTS deals in egocentric data might involve fraud. Vendors send a perfect sample. The final dataset is either full of recordings you already own or recordings already open-sourced on Hugging Face. Sellers can Rename, Trim, Crop, Re-encode. Checksums change. The content doesn’t. We just open-sourced Overlap. It indexes any OTS data and compares it against your inventory & public HuggingFace datasets. Ask sellers to share indexes first. Detect duplicates without ever seeing the real content. No GPUs needed. A normal MacBook can index 5k hrs data in an hour. Let’s change the industry: share index files instead of just 5mins samples. Stop paying for free or repeat data. Github - github.com/World-Archive/ove… GUI - overlap.worldarchive.co
1
2
2
18
6,521
World Archive retweeted
First ever in robotics to prevent fraud and duplicacy in data procurement. Check out github github.com/World-Archive/ove…
35% of OTS deals in egocentric data might involve fraud. Vendors send a perfect sample. The final dataset is either full of recordings you already own or recordings already open-sourced on Hugging Face. Sellers can Rename, Trim, Crop, Re-encode. Checksums change. The content doesn’t. We just open-sourced Overlap. It indexes any OTS data and compares it against your inventory & public HuggingFace datasets. Ask sellers to share indexes first. Detect duplicates without ever seeing the real content. No GPUs needed. A normal MacBook can index 5k hrs data in an hour. Let’s change the industry: share index files instead of just 5mins samples. Stop paying for free or repeat data. Github - github.com/World-Archive/ove… GUI - overlap.worldarchive.co
1
1
75
35% of OTS deals in egocentric data might involve fraud. Vendors send a perfect sample. The final dataset is either full of recordings you already own or recordings already open-sourced on Hugging Face. Sellers can Rename, Trim, Crop, Re-encode. Checksums change. The content doesn’t. We just open-sourced Overlap. It indexes any OTS data and compares it against your inventory & public HuggingFace datasets. Ask sellers to share indexes first. Detect duplicates without ever seeing the real content. No GPUs needed. A normal MacBook can index 5k hrs data in an hour. Let’s change the industry: share index files instead of just 5mins samples. Stop paying for free or repeat data. Github - github.com/World-Archive/ove… GUI - overlap.worldarchive.co
5
3
29
4,195
World Archive retweeted
Physical AI evals are starting to become a real category. Until recently, most VLA evaluation was basically: LIBERO → success rate → leaderboard. Now we're seeing a much broader stack emerge: • Allen AI — unified VLA eval across 18+ simulation benchmarks • LeRobot — one eval interface across multiple sim benchmarks • PhAIL — real robots + production metrics like throughput and failures • Robocurve — independent, real-world robot evaluation • RoboDojo — bringing sim + real-world evaluation together The interesting part isn't just more benchmarks. It's the move from: “Can the robot complete this task?” to: “How reliable, fast and general is this system in the physical world?” I think independent physical AI evals are going to become increasingly important as robot models start looking more and more similar on demos.
9
20
1
212
32,698
Excited to interview with @hf0 and share what we are building at @WorldArchive. Building pipelines to make robots learn better.
1
2
6
266
Scaling laws in robotics are real. But monocular ego alone is hitting diminishing returns. The next year belongs to rich DataOps — dense labels, multimodal capture, rights-clear delivery. New essay ↓ worldarchiveai.substack.com/… #PhysicalAI #robotics
1
7
158
Packaging lines, catering kitchens, auto body shops, artisanal work — egocentric manipulation data with "dense labels", not just raw hours. Load the sample ↓ huggingface.co/datasets/Worl…
2
5
121