Bo Wu

Data in the Wild, ep. 004. Task context • Clean several pairs of shoes, then arrange by type + use frequency • 11 minutes • A home in India Capture • Handheld monocular Realme RMX3867 • 1280×720, H.264 Embodied behavior • Repeated handling across shoes, cloth, cleaner + storage • Actions: wipe_shoe, rotate_shoe, apply_cleaner, change_pair, carry_shoes • Challenges: object identity, state tracking, tool changes + long occlusions Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets. Long-horizon tasks require more than recognizing an action frame by frame. Models need to track which objects they’ve interacted with, how their state has changed, and what still needs to happen.

Data in the Wild, ep. 005. Real-world tasks rarely happen from a fixed viewpoint. This one captures manipulation in a constrained space, with frequent camera movement, occlusion, and re-entry as the task progresses. Task context • Clean one area of a car interior + return supplies • 15 minutes • A parked car in residential United Kingdom Capture • Handheld monocular Infinix X665B • 1280×720, H.264 Embodied behavior • Cloth, bucket, dashboard, controls, storage, console + seats • Actions: rinse_cloth, wring_cloth, wipe_dashboard, wipe_control, open_storage, wipe_console, wipe_seat_gap, wipe_sill • Challenges: constrained reach, shifting viewpoints + repeated exits Video sourced from Numo, which collects real-world data that Poseidon then refines into training-ready datasets.

See you at @corl_conf 2026. Our paper, "When Crowdsourced Egocentric Video Goes Wrong: Failure Modes, Metadata, and Quality Control," was accepted to the "Oops, I Erred" workshop at CoRL, a leading robotics research conference. If you'll be in Austin, say hi: @sebliu + team will present our research on a critical challenge for Physical AI: how to ensure crowdsourced, real-world video data is reliable and useful for training robots. corl.org/

Our biggest speech data release yet: Today, we’re introducing our Indic Speech Release on @huggingface: 3,845+ hours of transcribed speech across Hindi, Telugu, Tamil, and Bengali. • 353K+ recordings • Longer, real-world speech samples • Speaker + demographic metadata • Per-example validation signals Built to help teams train, evaluate, and find the failures that short-clip benchmarks tend to hide.

Data in the Wild, ep. 002. Task context •- Clean a small aquarium + replace water •- 10 minutes •- A home in the Philippines Capture •- Handheld monocular Samsung SM-A135F •- 1280×720, H.264 Embodied behavior •- Two-handed handling across glass, water, tools + small loose objects •- Actions: remove_decor, transfer_water, rinse_vessel, replace_decor, fit_lid, wipe_surface, return_container •- Challenges: transparent water, reflective glass, wet-tool handling + long-horizon sequencing Video collected from Numo, our data collection app that sources real-world data. Poseidon then refines these captures for model training, with a focus on long-tail tasks.

Data in the Wild, ep. 001. The data needed to train robotics is rare, unscripted, and full of long-tail behaviors. Here’s what that looks like in practice. Task context: •- Brush the pet, wipe its paws & body, collect loose fur •- 11 minutes •- A home in the Philippines Capture: •- Handheld monocular iPhone •- 1920×1080 H.264 Embodied behavior: •- One hand stabilizes or repositions the dog while the other brushes •- Actions: stabilize_pet, brush_coat, reposition_pet, set_down_brush •- Challenge: Moving occlusion, careful hand manipulation, unpredictable animal motion

Physical AI has a data problem. To operate in the real world, models need to learn from human experience. The scale of what that demands is increasingly clear: ↳ $7T potential humanoid market by 2050 ↳ $11.9B → $49.8B AI training dataset market by 2031 ↳ ~100K hours of training data today, with tens of millions being targeted Meeting that scale will require vastly more training data than exists today. It’s where we’re increasingly focused at Poseidon.