exp: 轮次25——实时节奏(commit3)准确率：候选平均29.7%/冻结编码器34.4%/commit4 28.9%，均低于基线36.3%；效率：ORT CPU推理3.9ms+gm_groups，单次规划~32ms，50Hz异步可行

2026-10-20 12:49:37 +08:00
parent 79519b042c
commit 7e99b333ea
2 changed files with 27 additions and 2 deletions
@@ -237,6 +237,11 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
                score = torch.where(last_ok[:, None], -(d2 * w).sum(-1), torch.zeros_like(d2[..., 0]))
            order = torch.sort(score, dim=-1, descending=True, stable=True).indices
            pick = order[:, 0]
+            if select == 'average':   # 2026-10-20：候选平均（平滑误差）——所有候选置为 K 条动作的均值
+                inc = inc.mean(1, keepdim=True).expand_as(inc).contiguous()
+                K_eff = 1
+            else:
+                K_eff = K
            commit = inc[ar, pick, r:r + commit_n]
            if collect is not None:   # 反馈蒸馏：收集与部署完全一致的 (特征, 复合动作=模型输出+目标反馈)
                collect.append((feat.detach().float().cpu(), inc[ar, pick].detach().float().cpu()))
@@ -244,7 +249,7 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
            if safety and safety_partial:   # 2026-10-20：允许「部分安全前缀」——执行到第一个不安全步为止，而不是全拒刹停
                maxlen = torch.zeros(B, dtype=torch.long, device=device)
                chosen = torch.zeros(B, commit_n, 6, device=device); chosen_k = pick.clone()
-                for kk in range(K):
+                for kk in range(K_eff):
                    ck = inc[ar, order[:, kk], r:r + commit_n]
                    pts = qe[:, None] + torch.cumsum(ck, 1)
                    mids = torch.cat([qe[:, None], pts[:, :-1]], 1) + 0.5 * ck
@@ -263,7 +268,7 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
            elif safety:
                ok_any = torch.zeros(B, dtype=torch.bool, device=device)
                chosen = torch.zeros(B, commit_n, 6, device=device); chosen_k = pick.clone()
-                for kk in range(K):
+                for kk in range(K_eff):
                    ck = inc[ar, order[:, kk], r:r + commit_n]
                    pts = qe[:, None] + torch.cumsum(ck, 1)
                    mids = torch.cat([qe[:, None], pts[:, :-1]], 1) + 0.5 * ck