Demo risk: Static camera angles hide model brittleness. Real test = same task, different viewpoint.
@axisrobotics addresses this by replaying community demos with randomized cameras, lighting, textures, and object positions. Converts single recordings into varied training sets.
Dataset V1 results: Continued pretraining lifted π0.5 success rate on LIBERO-Plus from 83.9% → 88.8% (+4.9pp).
Key value prop = data engine that injects distributional variance into crowd-sourced demonstrations. Thesis: Robots trained on varied replays generalize better to out-of-distribution conditions.
Relevant for robotics infrastructure plays and foundation model scalability in physical AI.
@axisrobotics addresses this by replaying community demos with randomized cameras, lighting, textures, and object positions. Converts single recordings into varied training sets.
Dataset V1 results: Continued pretraining lifted π0.5 success rate on LIBERO-Plus from 83.9% → 88.8% (+4.9pp).
Key value prop = data engine that injects distributional variance into crowd-sourced demonstrations. Thesis: Robots trained on varied replays generalize better to out-of-distribution conditions.
Relevant for robotics infrastructure plays and foundation model scalability in physical AI.