Widespread broken code and training data in academic robotics projects
carlosdponx · x · 2026-11-05
The author isn't surprised academic robotics projects stall at 80-90% completion given lab incentives, but is genuinely surprised how many ship glaring code errors from copy-pasted, copy-of-a-copy codebases. He also reports tons of outright broken data used to train policies, and fixing that materially improves quality. Training code being researchers' core output, he wonders why it isn't deeply inspected.
More from Embodied
- A 1X humanoid robot quietly stands in an apartment, teased for upcoming Core Memory feature — kyliebytes · 2026-11-05
- Bambu H3C draws up to 1850W, needs dedicated 20A outlet or it will trip breakers — RexDouglass · 2026-11-05
- Robotics sim frontier names from World Labs, NVIDIA, PI to meet in SF on sim2real gap — jia_seed · 2026-11-05
- Robotics RL still on Isaac Lab PhysX despite mjwarp's 2x+ training speedup — carlosdponx · 2026-11-05
- General Intuition: game clips (1.5T action tokens/month) train one model to control 10+ robots — AI Engineer · 2026-11-05
- CoRL paper GUILD: actively guiding robot data collection toward useful auxiliary tasks — RobobertoMM · 2026-11-05