comma.ai breakdown: BF16 output layers created 0.125 m/s speed staircases in driving plans
yassineyousfi_ · x · 2026-09-21
comma.ai's blog post reviews ML bugs that broke its self-driving stack. One key bug: running the plan output head of a DiT world model in BF16 meant speeds near 30 m/s were representable only in 0.125 m/s steps, and the acceleration computation (differencing nearby predictions over a short Δt) amplified the rounding error. The fix: promote the plan head and its inputs to FP32 before running—casting the finished plan to FP32 just yields the staircase with more decimals. The post frames these bugs against the Linux kernel's 0.5 bugs per 1,000 LOC and how to catch such issues earlier.
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12