Skip to content
MouseMouse

The advisor's "runtime five-stage state machine + per-stage reference poses + RL residual" appeared in none of the three papers it cited - reading the originals changed the plan and downgraded two widely repeated industry claims

A lesson the Coach cites as advisor-paraphrase-vs-paper.
ReplicatedExperiment method

Read the primary source behind any piece of advice before adopting its architecture; record where the paraphrase and the original differ, downgrade claims the originals do not support to speculation, and adopt what the verified sources actually share.

Symptom

After the violent first real run, an advisor proposed re-architecting recovery as a runtime staged state machine with reference poses and an RL residual, citing HoST, HumanUP and StableMimic.

Context

The three papers were read in full on 2026-08-09 and tabulated (deployment form, what the "stages" really are, hard constraints, references). HoST: one end-to-end policy, height-gated rewards in training, action anchored as q + beta*a with a beta curriculum. HumanUP: two training stages with the same observation/action; the vendor's three-stage state machine is the baseline it beats (41.7% vs 78.3%); Stage II tracks an 8x slowed Stage I trajectory (4x too violent, 10x does not converge). StableMimic: a learned soft gate. On 08-10 more sources were checked the same way: the Agility page does not say Digit's self-righting was learned in simulation (only step recovery is stated as RL), so the claim was downgraded to speculation; HoST's support for the 12-DoF armless Mini Pi exists in its code repository, not in the paper text.

Change

Adopted only what the originals share: a hard action bound or anchored action space, strong smoothing including a second-difference term, a slowed hidden reference, heavy DR and real fallen states. The staged runtime state machine was not adopted.

Outcome

The next re-rooting candidates came straight from the verified material, and the beta-anchored action space (HoST, with the Mini Pi configuration as the nearest real-robot precedent) became V2, the lineage that later stood up on hardware.

Mechanism

A paraphrase compresses a paper into the advisor's own architecture; only the original shows what was actually deployed, what was a baseline, and which numbers came with which ablation.

Applies when

  • an advisor, agent or summary proposes an architecture with citations
  • an industry claim ("X learned it in sim") is about to justify a design
  • several papers are cited for one combined recipe
“⇒ **顾问的核心形态"runtime 五阶段状态机 + 每阶段参考姿态 + RL residual"在三篇引文 里均不存在**,其中 HumanUP 还点名 state machine 是局限。”
git:Lucen-recovery@origin/recovery:train/RECOVERY_V0_SPEC.md § §26 三篇引文精读判决(2026-08-09 全文核对;顾问转述与原文有出入)

Principles that cite it