Poseidon (@psdnai) header
Poseidon (@psdnai) profile picture

Real-world data for physical AI

Joined July 2025
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
3
5
3
25
3,129
Failed segment ≠ failed recording. Failure-aware QC can salvage usable footage, without discarding entire videos. In our #CoRL2026 paper, we audited 500 hours of footage to understand the failures that affect Physical AI training data. Full paper: openreview.net/pdf?id=rLGRN1…
3
1
1
19
1,681
Key findings: • 81% of clips with certain annotated failures had issues covering no more than half their duration • Targeted trimming preserved up to 91.7% of footage otherwise lost to whole-clip rejection • Standard automated quality checks showed no clear improvement over a flag-all baseline in our tests TL;DR: a flawed video isn't necessarily unusable, and rejecting an entire recording because of localized failures sacrifices valuable training data. Quality control should be about recovering usable training data, not just rejecting bad training data ↴ huggingface.co/spaces/psdn-a…
5
109
Poseidon retweeted
so excited to share this! i led the ASR research and experiments behind this work. 10,000 hours of Bengali speech data, ~32% lower CER on Whisper large v3 :) grateful for Sandeep Chinchali’s mentorship!
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
2
3
360
Better data beats bigger models. That's what the latest @psdnai case study proves. 10K hours of Bengali speech, built for how people actually speak: ▪️ Regional dialects + code-switching ▪️ Verified native speakers, linguists in the loop ▪️ Whisper Large V3 fine-tuned → ~32% lower CER ▪️ Beat India’s top sovereign API Proud to have worked alongside @iamrashminagpal @jdubpark @ZK_pawgchain @snehasanks on this release! Full case study ↓ xitter.cf/psdnai/status/21227556…
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
1
12
423
Poseidon retweeted
Very proud to have shipped this case study, alongside legends: @iamrashminagpal @jdubpark @mskristagambrel @ZK_pawgchain @snehasanks @nickipp12 onward ⚔️ psdn.ai/blog/case-study-10k-…
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
1
1
5
212
Truth.
Replying to @chris_j_paxton
The problem with robotics is almost everything is an edge case
2
11
1,490
Most data collection starts with what’s easy to collect. At @PsdnAI we start with what models get wrong. We care about how data improves the models. (1) Create a holistic benchmark (2) Find where models & APIs fail (3) Run collection campaigns across high-value channels for that data (4) QC and annotate (5) Fine-tune, improve, & benchmark OSS models on that data -> loop We’re now taking the same loop to Physical AI. We care about how ego video, audio, and tactile data improve models and robots in the wild. DM me for more info on how you can get involved and benefit from this loop!
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
2
7
347
Mentally, we're still at @corl_conf. Our highlights: ▸ Robots are more accessible than ever. Deployment is the next frontier ▸ Real-world deployment data is becoming robotics' most valuable asset ▸ @RussTedrake reinforced the importance of deployment data, showcasing robotics on a Toyota factory line ▸ Dexterity, tactile learning, and real2sim2real are gaining momentum ▸ Human video is increasingly being converted into executable robot actions More from @jdubpark ↓
CoRL takeaway: Robots are now fully accessible... deploying is next. With off-the-shelf hardware and agentic robotics (Astra), you can put together a working robot. It will be slow and make mistakes. But this turns the question from “can anyone build an e2e robot pipeline?” to “how do you improve one?” The race is now on to get them working in the field. So the visible moat is deployment knowledge. WAMs/VLAs, sim, hardware, policies, etc. are all being commoditized. And every architecture and method chases the same goal: robots in prod at scale. Ofc, to get there you need a closed feedback loop that includes deployment, like Optimus in the Gigafactory or Amazon FAR in warehouses. Deployment is the back-propagation for robotics. Despite the accessibility, the gap between research, hardware, and real-world usage remains wide. Walden Robotics founder @RussTedrake's keynote impressed many with “deployment data is the key” and a demo on a Toyota factory line. Tbh, that’s common knowledge. But in our minds, the path from software to robots working in production lines is still very murky. And this conference is a great wake-up call for roboticists and companies to be more practical -- get your hands dirty in factories and homes, break things and learn to recover in prod. Perfection matters more than generalization in most deployments - failure is unacceptable. Agentic robotics is emerging fast! Astra and Fable can control robots with no extra training decently (pre-trained on ego videos :p). It’s slow for now, so most people treat it as long-horizon planning. The stack is settling into System 2 (what to do), System 1 (how to do it - agentic robotics), and System 0 (real-time actions). More notes: Progress to a functioning robot in the wild depends on connecting three capabilities: - Learning from human demonstrations - Reacting to contact and the environment with hands and whole body - Retargeting data & skills across different hands and bodies Human video is becoming a source of executable demonstrations, not just for training for generalization. EgoScale, Do as I Do, and Video2Sim2Real all convert video into actions a robot can execute. Lots of imitation learning alongside RL (though it's doubtful whether it's better than RL for prod). Deployment data is the gold standard, since everything in robotics exists to make deployment work. Walden’s 4.7B “Sage” model trained on 100k hours from a Toyota line and performs very well on that task (unsure for other tasks). Teleop interventions catch failures and record them to improve the model. Synthetic data is growing too, ultimately for real2sim2real: bring reality into sim, train and test at scale, transfer back. GRAIL, Play2Perfect, etc. Dexterity got a lot of attention. T-Rex used 100 hours of tactile data to sharply raise task success. Some papers drop vision and rely on touch alone... but again, why? Feels very detached from practicality. Unlike Walden, tactile research is still far from industry deployment.
1
2
21
1,946
Fine-tuning Whisper on that data: • Reduced CER by ~32% on held-out test data • Outperformed every open-source model we tested • Overtook India's top sovereign API • Closed most of the gap to leading commercial APIs
1
1
145
Beyond Bengali, it's evidence that the next wave of AI performance will come from data that's representative, difficult, and grounded in how people actually communicate. Read the full case study: psdn.ai/blog/case-study-10k-…
2
121
1/ Fun #CoRL2026 dinner hosted by @GradientVC & @goaly_ai in Austin earlier this week, with a great crowd of robotics builders! Hot debates: can Astra and Fable run robots on their own or do we still need robot FMs? And does more video actually make for better data? 🧵 thanks for everyone who joined us! @AdemiAdeniji (@GroundedSI), @drxcliu (@DeepReach_AI ), @itsghvr (@MeckaAI), @jamesw_stevens (@enact_labs), @krisPsdn(@psdnai), @kun_h____ (@DynaRobotics), @kylerduffy, @markgrinev(@MeckaAI), @NickelReady (@DatoricAI), @Parvkpr(@genrobotics_ai), @praveenvnktsh (@Uber), @DanielXieee(@origami_robo), @sebliu(@psdnai), @Anciukevicius(@Waymo), @vincentjliu(@GroundedSI), @liuying04(@scale_AI)
3
1
1
43
2,849
Someone tried to sell us our own AI training data. A third-party provider included data we'd already collected in a sample dataset sent to us for evaluation. Our team caught it immediately. How often is training data being resold without the buyer knowing?
13
2
1
51
9,103
Meet Rashmi, one of the driving forces behind Poseidon's research on quality control in crowdsourced egocentric video. Rashmi spearheaded much of the work behind our paper, "When Crowdsourced Egocentric Video Goes Wrong," presented today at #CoRL2026 in Austin. If you're still in Austin, drop us a line and learn what we've uncovered about the challenges of collecting reliable data for Physical AI.
2
2
1
28
2,400
Poseidon retweeted
Very high signal contribution to the robotics training space from Poseidon’s bleeding edge research! Open for anyone to use. Happy training and evaluating 💪
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
3
4
24
1,257
We will open source more data and research as we expand our capacity. For researchers at CoRL, reach out for various data! Happy to collaborate on improving models with globally diverse datasets
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
1
4
14
553
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
3
7
3
40
6,234
Explore the data, inspect the failure modes, and see how crowdsourced video can impact Physical AI: huggingface.co/spaces/psdn-a…
10
397
The team is reporting live from #CoRL2026 ↓
The theme is "dexterity" at CoRL. Most booths are humanoid hands and robot arms! Cool hands from @origami_robo. @PsdnAI is researching visuo-tactile WAMs with own video and tactile data. Reach out!
4
31
2,300