feat(abs-target): 方案 B 初版——Y 改 16 步绝对目标角（/180），训练/评估/rollout/ckpt 适配；修复前缀噪声增广对绝对目标的标签污染（eps 直接加到绝对位置=18°）；实测一步生成误差 31°、闭环 0%，此配方不成立

ml/sim/closed_loop.py

@@ -153,7 +153,10 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
     B = q0.shape[0]
     step = torch.tensor(STEP, device=device)
     act_scale = oracle.act_scale if oracle is not None else float(getattr(model, 'act_scale', 1.0))
-    astep = step * act_scale                         # 动作反归一化：每单位 act 对应的度数
+    # 方案 B（2026-10-19）：abs_target 模型输出 16 步绝对目标角（/180 归一化），执行时直接朝目标走；
+    # oracle / 旧 ckpt 仍是增量口径
+    abs_t = oracle is None and bool(getattr(model, 'abs_target', False))
+    astep = torch.full((6,), 180.0, device=device) if abs_t else step * act_scale
     has_ref = torch.isfinite(ex)
     limit = time_limit(ex, max_sec)
     nmax = int(math.ceil(limit.max().item() / DT))
@@ -177,8 +180,13 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
         tnow = it * DT
         if (queue.shape[1] <= 3) if replan == 'r3' else (queue.shape[1] == 0):
             r = queue.shape[1]
-            pre = torch.zeros(B, 3, 6, device=device); pre[:, :r] = queue[:, :r]
-            qe = q + queue.sum(1)
+            pre = torch.zeros(B, 3, 6, device=device)
+            if abs_t and r > 0:                          # queue 存绝对目标：前缀特征仍用增量（度）
+                prevq = torch.cat([q[:, None], queue[:, :r - 1]], 1)
+                pre[:, :r] = queue[:, :r] - prevq
+            else:
+                pre[:, :r] = queue[:, :r]
+            qe = (queue[:, -1] if r > 0 else q) if abs_t else q + queue.sum(1)
             e, oe = goal_err(kin, qe, T, R)
             near = (e < 5) & ((R.abs().sum(-1) == 0) | (oe < 5))
             subs, nsub, sub_prev, sub_t = sub_advance(kin, subs, nsub, sub_prev, sub_t, qe, tnow, sub_rule)
@@ -189,7 +197,7 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
             else:
                 z = torch.randn(B * K, H, 6, device=device, generator=g) * z_scale
                 act, val, sub = model.sample_k(feat, K, z, steps=steps)
-            inc = act.float().clamp(-1, 1) * astep
+            inc = act.float().clamp(-1, 1) * astep   # 增量口径=步长增量；B 口径=绝对目标角
             if cold:
                 need = sub[:, 12] > 0.5
                 subs = q[:, None] + sub[:, :12].float().view(B, 2, 6) * 180
@@ -203,7 +211,7 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                     nsub = torch.where(need, torch.full_like(nsub, 2), torch.zeros_like(nsub))
                 if not use_sub: nsub = torch.zeros_like(nsub)
                 sub_t.fill_(tnow); sub_prev = q.clone()
-            fut = qe[:, None, None] + torch.cumsum(inc[:, :, r:], 2)          # [B,K,H-r,6]
+            fut = inc[:, :, r:] if abs_t else qe[:, None, None] + torch.cumsum(inc[:, :, r:], 2)   # [B,K,H-r,6] 绝对位置
             if select == 'first':
                 score = torch.zeros(B, K, device=device); score[:, 0] = 1
             elif cold or select == 'value':
@@ -231,8 +239,12 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                 chosen = torch.zeros(B, 3, 6, device=device); chosen_k = pick.clone()
                 for kk in range(K):
                     ck = inc[ar, order[:, kk], r:r + 3]
-                    pts = qe[:, None] + torch.cumsum(ck, 1)
-                    mids = torch.cat([qe[:, None], pts[:, :-1]], 1) + 0.5 * ck
+                    if abs_t:
+                        pts = ck
+                        mids = 0.5 * (torch.cat([qe[:, None], ck[:, :-1]], 1) + ck)
+                    else:
+                        pts = qe[:, None] + torch.cumsum(ck, 1)
+                        mids = torch.cat([qe[:, None], pts[:, :-1]], 1) + 0.5 * ck
                     allq = torch.cat([pts, mids], 1).reshape(-1, 6)
                     okk = (kin.gaps(allq, 0.5)['gm'].min(-1).values > 0).view(B, 6).all(-1)
                     take = okk & ~ok_any
@@ -242,6 +254,7 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                 pick = chosen_k; committed = ok_any
             if near.any():
                 ref = dls_steps(kin, qe, T, R)
+                if abs_t: ref = qe[:, None] + torch.cumsum(ref, 1)   # 增量 → 绝对目标
                 commit = torch.where(near[:, None, None], ref, commit)
                 dls_it = torch.where(near & (dls_it < 0), torch.full_like(dls_it, it), dls_it)
                 committed = committed & ~near
@@ -250,9 +263,12 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
             queue = torch.cat([queue, commit], 1)
             cold = False
         d = queue[:, 0]; queue = queue[:, 1:]
-        d = torch.where(done[:, None], torch.zeros_like(d), d)
-        qp = q; q = kin.clamp_box(q + d)
-        travel += d.abs().sum(-1)
+        if abs_t:                                     # 绝对目标：直接走到目标；到位的题原地冻结
+            qp = q; q = kin.clamp_box(torch.where(done[:, None], q, d)); travel += (q - qp).abs().sum(-1)
+        else:
+            d = torch.where(done[:, None], torch.zeros_like(d), d)
+            qp = q; q = kin.clamp_box(q + d)
+            travel += d.abs().sum(-1)
         j1_acc = j1_acc + d[:, 0]; j1_max = torch.maximum(j1_max, j1_acc); j1_min = torch.minimum(j1_min, j1_acc)
         j1_first = torch.where((j1_first == 0) & (j1_acc.abs() >= FIRST_DEG), torch.sign(j1_acc), j1_first)
         if it + 1 == n_start: moved_start = travel.clone()