Tatul Danielyan (@daniel7an) header
Tatul Danielyan (@daniel7an) profile picture
Joined February 2025
Photo by Tatul Danielyan (@daniel7an)Photo by Tatul Danielyan (@daniel7an)Photo by Tatul Danielyan (@daniel7an)
Proud of the team for shipping these open models 🚀
🇦🇲 3 open Armenian AI models from Metric AI Lab: 🎙️ OhanASR: speech → text, incl. streaming 🔊 OhanTTS: text → speech + voice cloning 🧠 Ayb: LLM + vision (0.8B/4B/9B) Built for the community. 🔗 metricailab.com/blog/ 🤗 huggingface.co/Metric-AI #ArmenianAI #OpenSource
1
9
55
🏆 Announcing the RoboEval Challenge winners at #CoRL2026! 🥇 @IliaLarchenko — 0.988 🥈 Metric AI Lab (@daniel7an, @HrantDavtian) — 0.982 🥉 LocalModels — 0.978 Across 8 bimanual manipulation tasks, we evaluated not just task success, but also safety, efficiency, and coordination. Huge congratulations to our winners! 🎉🤖
1
3
1
9
828
🥈 We took 2nd place in the RoboEval Diagnostic Bimanual Challenge at #CoRL2026! We combined vision based object pose estimation with scripted controllers to complete 8 bimanual manipulation tasks, using just 14.8 GPU-hours of training. Presented by @HrantDavtian. Thanks @YiruHelenWang for organizing! Code + weights: github.com/Metric-AI-Lab/rob…
3
1
1
17
888
Had a great time giving a short talk on retrieval for legal question answering in low-resource languages at this BoF. Thanks @JinaAI_ for organizing! I'm also presenting a poster tomorrow (Mon), 11:00-12:30, Hall H, #949. Come say hi if you're at #EMNLP2026 🇭🇺
Our Birds of a Feather session on embeddings, rerankers & small language models for better search went off — full house at EMNLP 2026! Shoutout to everyone who made it. Enjoy the rest of the conf in Budapest! docs.google.com/presentation…
1
8
471
Tried real-to-sim on our SO-101. Gave Astra top + wrist video and the robot actions. It built the scene, calibrated the cameras, and replayed the task in MuJoCo + Blender.
GPT-6 does real2sim end to end from robot demonstrations. I gave it only multi-view RGB and robot actions, and one simple prompt. It calibrated the cameras, built the object assets, performed physics system-ID, ran MuJoCo, and rendered in Blender.
2
8
1,117
Tatul Danielyan retweeted
Gave GPT-6 Astra (low) and Claude Fable 5.1 (low) the same SO-101. Both models got a calibration file, a camera/state screenshot script, and: - “control the lerobot so101 robot arm and perform a pick and place task using commands and visual camera as input” > GPT-6 Astra (low): done in 9:45 > Claude Fable 5.1 (low): stopped at 15:10, not done
1
10
493
Tatul Danielyan retweeted
controversial take but i feel system design should be learnt experiencing your naive approach fail than the other way around. theory can only take you so far
24
3
3
192
7,670
Tatul Danielyan retweeted
The inspiration you need
62
347
58
2,727
144,693