exp+fix: commit2 增益重扫（gain2.2 到达 50.8%，n=512）+ 浏览器安全层三点修复（限位夹紧/部分前缀/离散认证）
- gain 1.8/2.2/2.5/2.8 = 50.4/52.0/49.6/47.7%（n=256），2.2 最佳；n=512 复核 50.8%/1.4% 碰撞 - 积分反馈与反馈隔次复用均为负结论（−4.0/−5.1 pp） - JS：tick/候选投影夹 LIM（原越限后 inBox 永久拒绝→刹停）；safety_partial 移植；segCert 开关（默认严格） - bench nc10 --seg 0：刹车率 79.5→48.1%、planP50 56.1→33.5ms、ok 0→5/10
This commit is contained in:
+63
-18
@@ -138,7 +138,19 @@ function refineStep(S, q, T, Rt, maxScale) {
|
||||
return r.concat([0.5 * (M[7] - M[5]) * 100, 0.5 * (M[2] - M[6]) * 100, 0.5 * (M[3] - M[1]) * 100]);
|
||||
};
|
||||
const e0 = e(q), m = e0.length, h = 1e-3, J = [];
|
||||
for (let j = 0; j < 6; j++) { const qa = q.slice(); qa[j] += h; const e1 = e(qa); J.push(e1.map((v, i) => -(v - e0[i]) / h)); }
|
||||
if (S.jac) {
|
||||
/* 解析几何雅可比（与 ml/sim/closed_loop.py 的 dls_steps(analytic=True) 同口径）：
|
||||
有限差分在 float32/双精度下都会给反馈方向注入 ~1–4% 噪声，闭环里会被慢变误差放大。
|
||||
J[j][i] = −∂err_i/∂q_j：位置 err=T−p → +Jv；姿态 err=100·rotvec → +100·Jw。 */
|
||||
const jv = S.jac(q);
|
||||
for (let j = 0; j < 6; j++) {
|
||||
const col = [jv.Jv[0][j], jv.Jv[1][j], jv.Jv[2][j]];
|
||||
if (m === 6) col.push(100 * jv.Jw[0][j], 100 * jv.Jw[1][j], 100 * jv.Jw[2][j]);
|
||||
J.push(col);
|
||||
}
|
||||
} else {
|
||||
for (let j = 0; j < 6; j++) { const qa = q.slice(); qa[j] += h; const e1 = e(qa); J.push(e1.map((v, i) => -(v - e0[i]) / h)); }
|
||||
}
|
||||
// Δq = Jᵀ (J Jᵀ + λ²I)⁻¹ e，m ≤ 6
|
||||
const lam = 0.5, A = [];
|
||||
for (let i = 0; i < m; i++) { A.push([]); for (let k = 0; k < m; k++) { let s = i === k ? lam * lam : 0; for (let j = 0; j < 6; j++) s += J[j][i] * J[j][k]; A[i].push(s); } }
|
||||
@@ -179,14 +191,14 @@ class Runner {
|
||||
this.o = Object.assign({ K: 8, commit: 3, margin: 5, /* 2026-10-18：运行时投影余量保持 5 mm（真机护栏，不迁就数据分布）；数据侧另行过滤认证间隙 < 3 mm 的老师轨迹，APG 训练余量 3 mm */
|
||||
projMax: 0.5, refineMm: 5, arriveMm: 1, arriveDeg: 0.5,
|
||||
stuckPlans: 5, lockSec: 1.0, subReachDeg: 5, subTimeout: 1.5, subRule: 'cross', subLeadSec: 0.02, seed: 12345, lookahead: true,
|
||||
actScale: 1, certBudget: 40, safetyMs: 30, firstMs: 45, fbGain: 0 }, opts || {});
|
||||
actScale: 1, certBudget: 40, safetyMs: 30, firstMs: 45, fbGain: 0, segCert: true }, opts || {});
|
||||
/* 2026-10-19：actScale = 模型 meta.act_scale（动作标签按 act_scale·STEP 归一化，旧模型 1）；增量 = clamp(act, ±1)·actScale·STEP，
|
||||
与 ml/sim/closed_loop.py 同口径。subRule：'cross'（默认，越过顶点即切，与训练标签同口径；见 _subAdvance）/ 'timeout'（旧：< subReachDeg 或 subTimeout 超时）。 */ // 每条候选最多 40 次距离查询；安全层总预算 30 ms，第 1 条候选可用到 45 ms（超出按刹停处理）
|
||||
this.cert = makeCert(S, { leverTable: this.o.leverTable, cap: 20 });
|
||||
this.rng = makeRng(this.o.seed);
|
||||
/* 2026-10-18 第六轮（审查 B-1/B-2）：耗时拆开记——inferMs 主推理、lookaheadMs 前瞻（第二次推理 + 8 份特征）、safetyMs 只含认证、
|
||||
planMs 整次规划（R9 口径：必须 < 60 ms 周期）；certQ 每次规划新增的距离查询数，certHit 记忆命中数，certWhy 认证失败原因计数 */
|
||||
this.stats = { plans: 0, inferMs: [], lookaheadMs: [], safetyMs: [], planMs: [], planWarmMs: [], planColdMs: [], certQ: [], certHit: 0, certWhy: {}, brakes: 0, refine: 0, lookahead: 0, locks: 0, projected: 0 };
|
||||
this.stats = { plans: 0, inferMs: [], lookaheadMs: [], safetyMs: [], planMs: [], planWarmMs: [], planColdMs: [], certQ: [], certHit: 0, certWhy: {}, brakes: 0, refine: 0, lookahead: 0, locks: 0, projected: 0, partial: 0 };
|
||||
this.reset(null);
|
||||
}
|
||||
reset(q0) {
|
||||
@@ -213,7 +225,10 @@ class Runner {
|
||||
if (!this.q) return null;
|
||||
const d = this.queue.shift();
|
||||
this.qPrev = this.q.slice();
|
||||
if (d) this.q = this.q.map((v, j) => v + d[j]);
|
||||
if (d) this.q = this.q.map((v, j) => { // 与 ml/sim/closed_loop.py 的 clamp_box 同口径：限位轴执行时夹到 LIM
|
||||
const id = this.S.ORDER[j], L = this.S.LIM[id], x = v + d[j];
|
||||
return L.cont ? x : Math.max(L.lo, Math.min(L.hi, x));
|
||||
});
|
||||
this.t += DT;
|
||||
if (this.goal && !this.queue.length) {
|
||||
const { e, oe } = this.err(this.q);
|
||||
@@ -250,16 +265,39 @@ class Runner {
|
||||
}
|
||||
if (adv) { this.subPrev = this.subs.shift(); this.subT = this.t; }
|
||||
}
|
||||
_certCommit(qe, steps) { // 认证：提交步 + 刹停段（先用离散判定快速剔除，再做距离下界认证）
|
||||
const pts = [qe.slice()]; let q = qe.slice();
|
||||
for (const d of steps) { q = q.map((v, j) => v + d[j]); pts.push(q.slice()); }
|
||||
for (let i = 1; i < pts.length; i++) if (!this.S.collideFree(pts[i])) return { ok: false, tail: [] };
|
||||
const tail = brakeSteps(steps.length ? steps[steps.length - 1] : [0, 0, 0, 0, 0, 0]);
|
||||
/* 刹停段各步同向等比减速 → 关节空间里是一条直线，认证终点一段即可 */
|
||||
if (tail.length) { for (const d of tail) q = q.map((v, j) => v + d[j]); pts.push(q.slice()); }
|
||||
const c = this.cert.path(pts, this.o.certBudget, this._until);
|
||||
if (!c.ok) this.stats.certWhy[c.why || 'collide'] = (this.stats.certWhy[c.why || 'collide'] || 0) + 1;
|
||||
return { ok: c.ok, tail };
|
||||
_certCommit(qe, steps, partial) { // 认证：提交步 + 刹停段（先用离散判定快速剔除，再做距离下界认证）
|
||||
/* partial：与 ml/sim/closed_loop.py 的 safety_partial 同口径——整段不过时依次缩短前缀，
|
||||
执行到第一个不安全步为止（而不是整段拒绝 → 刹停）。2026-10-20：浏览器 bench 刹车率 47% 的主因是整段拒绝。 */
|
||||
const tryPath = n => {
|
||||
const use = steps.slice(0, n);
|
||||
const pts = [qe.slice()]; let q = qe.slice();
|
||||
for (const d of use) { q = q.map((v, j) => v + d[j]); pts.push(q.slice()); }
|
||||
if (!this.o.segCert) {
|
||||
/* 离散判定（步点 + 中点，与 ml/sim/closed_loop.py 安全层同口径）：capsule 模型余量已含在 collideFree。
|
||||
连续区间认证（segCert）~40 ms/次，是 50 Hz 部署的主要瓶颈；关闭后 planP50 预计 <30 ms。 */
|
||||
let prev = qe.slice();
|
||||
for (const p of pts.slice(1)) {
|
||||
const mid = prev.map((v, j) => 0.5 * (v + p[j]));
|
||||
if (!this.S.collideFree(mid) || !this.S.collideFree(p)) return null;
|
||||
prev = p;
|
||||
}
|
||||
const tail = brakeSteps(use.length ? use[use.length - 1] : [0, 0, 0, 0, 0, 0]);
|
||||
let qq = q;
|
||||
if (tail.length) { for (const d of tail) qq = qq.map((v, j) => v + d[j]); if (!this.S.collideFree(qq)) return null; }
|
||||
return { ok: true, tail: tail, steps: use };
|
||||
}
|
||||
for (let i = 1; i < pts.length; i++) if (!this.S.collideFree(pts[i])) return null;
|
||||
const tail = brakeSteps(use.length ? use[use.length - 1] : [0, 0, 0, 0, 0, 0]);
|
||||
/* 刹停段各步同向等比减速 → 关节空间里是一条直线，认证终点一段即可 */
|
||||
let qq = q;
|
||||
if (tail.length) { for (const d of tail) qq = qq.map((v, j) => v + d[j]); pts.push(qq.slice()); }
|
||||
const c = this.cert.path(pts, this.o.certBudget, this._until);
|
||||
if (!c.ok) this.stats.certWhy[c.why || 'collide'] = (this.stats.certWhy[c.why || 'collide'] || 0) + 1;
|
||||
return { ok: c.ok, tail: tail, steps: use };
|
||||
};
|
||||
if (!partial) { const r = tryPath(steps.length); return r || { ok: false, tail: [] }; }
|
||||
for (let n = steps.length; n >= 1; n--) { const r = tryPath(n); if (r && r.ok) return r; }
|
||||
return { ok: false, tail: [] };
|
||||
}
|
||||
async plan() {
|
||||
if (!this.needPlan()) return;
|
||||
@@ -316,7 +354,9 @@ class Runner {
|
||||
const adj = [];
|
||||
for (const d of steps) {
|
||||
let qn = q.map((v, j) => v + d[j]);
|
||||
const qp = project(S, qn, o.margin, o.projMax, near);
|
||||
let qp = project(S, qn, o.margin, o.projMax, near);
|
||||
/* 限位轴先夹到 LIM（与执行 clamp_box 同口径），否则候选会被认证的 inBox 整段拒绝 → 永远刹停 */
|
||||
qp = qp.map((v, j) => { const L = S.LIM[S.ORDER[j]]; return L.cont ? v : Math.max(L.lo, Math.min(L.hi, v)); });
|
||||
if (qp !== qn) this.stats.projected++;
|
||||
adj.push(qp.map((v, j) => v - q[j])); q = qp;
|
||||
}
|
||||
@@ -366,16 +406,21 @@ class Runner {
|
||||
collideFree 1.5 ms、criticalPairs 3.1 ms，推理本身 1.5 ms——耗时在几何查询，不在模型。此前截止只在候选之间检查，
|
||||
单条候选最多 40 次查询（≈150 ms）不受限。现在墙钟截止下沉到认证内部：第 1 条候选截止 t1+firstMs，其余共享到 t1+safetyMs；
|
||||
到点判不过 → 沿上次已认证的刹停段停，安全性不变，只是可能多刹停。 */
|
||||
let picked = null;
|
||||
let picked = null, pickedSteps = null;
|
||||
for (const c of pool) {
|
||||
if (c !== pool[0] && now() - t1 > o.safetyMs) break;
|
||||
this._until = t1 + (c === pool[0] ? Math.max(o.firstMs, o.safetyMs) : o.safetyMs);
|
||||
const v = this._certCommit(qe, c.steps); if (v.ok) { picked = c; this.brakeTail = v.tail; break; }
|
||||
const v = this._certCommit(qe, c.steps, true);
|
||||
if (v.ok) { picked = c; pickedSteps = v.steps; this.brakeTail = v.tail; break; }
|
||||
}
|
||||
this._until = undefined;
|
||||
this.stats.safetyMs.push(now() - t1);
|
||||
if (picked) {
|
||||
this.queue.push(...picked.steps); this.lastPlan = [qe.slice()].concat(picked.fut); this.brakeRun = 0;
|
||||
if (pickedSteps.length < picked.steps.length) this.stats.partial++;
|
||||
this.queue.push(...pickedSteps);
|
||||
/* 部分前缀时实际执行步数 < commit，一致性对齐的偏移不再成立 → 置空（Python 侧同轮次 14 的部分前缀满分对齐留待后续统一） */
|
||||
this.lastPlan = pickedSteps.length === picked.steps.length ? [qe.slice()].concat(picked.fut) : null;
|
||||
this.brakeRun = 0;
|
||||
} else {
|
||||
this.stats.brakes++; this.brakeRun++;
|
||||
if (this.brakeTail.length) { this.queue.push(...this.brakeTail); this.brakeTail = []; }
|
||||
|
||||
Reference in New Issue
Block a user