Discrete policies preserve expert motion: jerk ratio 0.74 vs 0.07 for continuous
ChongZzZhang · x · 2026-10-21
A robotics imitation learning study found:
- Core finding: continuous policies tend to over-smooth their actions, while discrete ones preserve more of the expert's temporal variation.
- Quantified: single-step continuous policies have a median jerk ratio (predicted jerk ÷ expert jerk on the same chunk; 1 = expert-matching) of just 0.07, vs 0.74 for discrete heads.
- Scene changes: discrete heads also react more like the expert when the scene shifts.
Directly relevant to choosing action heads in embodied AI systems.
More from Embodied
- Walden Robotics CTO joins Long Horizon panel on large behavior models with Rodney Brooks — adnothing · 2026-10-21
- YC-backed Antropi Robotics emerges with autonomous CNC factories, already profitable in 6 months — ycombinator · 2026-10-21
- Why the $3,000+ iPhone Duo is sold out until January: one buyer's theory — RayFernando1337 · 2026-10-21
- Robotics startup launches Deployments to serve fine-tuned HF checkpoints to robots in real time — arnie_hacker · 2026-10-21
- Amazon FAR's FIBER uses action-aligned latent planning to let Unitree G1 open doors and carry boxes — ChongZzZhang · 2026-10-21
- Sentdex trains robots to go from lying down to a half kneel, eyeing a full stand — Sentdex · 2026-10-21