After months of work, team reconstructs 3D human-object motion from plain video
andrew_n_carr · x · 2026-09-18
Andrew Carr announces that after months of effort, his team's pipeline started working this week: capturing a person and the objects they interact with in 3D — motion for both plus approximate object shape — from regular video.
- Key insight: robot learning and skill analysis need the object's motion too, not just the human's (tool use, tennis swings, animation props, occupational therapy).
- Still imperfect, but good enough to demo.
More from Embodied
- 60Hz e-ink on a Framework Laptop 13, driven by Modos' Paper Dev Kit — tweetsatpreet · 2026-09-18
- YC F26's Dreamscale Labs Moves Robot AI Inference to the Cloud — ycombinator · 2026-09-18
- Figure CEO Brett Adcock Hypes a Robot Demo That Won Over a Skeptic — adcock_brett · 2026-09-18
- Claimed GPT-6 Astra rapidly adapts to new robot embodiment, learns cake-to-air-fryer task — aran_nayebi · 2026-09-18
- Flix: An Open-Source ESP32 Quadcopter You Can Build From Scratch in Under 2,000 Lines — blaizedsouza · 2026-09-18
- OK-Robot Does Useful Home Chores With Zero Extra Training, Sparking Startup Hype Debate — notmahi · 2026-09-18