exp(dagger-deploy): 轮次20——反馈蒸馏（17.7万样本+3000步微调）负结论：gain=0 时到达仅0.8%（反馈未内化），gain1.8 与基线持平（43.4 vs 43.8）

2026-10-20 08:42:29 +08:00
parent ff160f576e
commit bb919d33c9
4 changed files with 119 additions and 1 deletions
@@ -142,7 +142,7 @@ class TeacherOracle:
@torch.no_grad()
 def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None, return_traj=False, use_sub=True, sub_override=None,
            qp0=None, steps=1, z_scale=1.0, select='consist', oracle=None, replan='r3', sub_rule='cross',
-            fb_gain=0.0, fb_scale=0.0, fb_near=0.2, fb_fallback=0.0, safety_partial=0, commit_n=3):
+            fb_gain=0.0, fb_scale=0.0, fb_near=0.2, fb_fallback=0.0, safety_partial=0, commit_n=3, collect=None):
    """use_sub=False：不用子目标（sf=0）；sub_override=(subs [B,M,6] 绝对关节角, n [B] 个数 或 need [B] bool)：冷启动时用给定途经点代替子目标头；
    qp0 [B,6]：起点的上一帧（默认 = q0，即静止起步）。steps：MeanFlow 采样段数；z_scale：噪声缩放；
    select：consist（部署口径：冷启动价值前瞻、之后与上一条已提交计划的一致性）/ value（每次都价值前瞻）/ first（不选，取第 0 条）。
@@ -234,6 +234,8 @@ def rollout(model, kin, tasks, device, K=8, safety=False, seed=0, max_sec=None,
            order = torch.sort(score, dim=-1, descending=True, stable=True).indices
            pick = order[:, 0]
            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()))
            committed = torch.ones(B, dtype=torch.bool, device=device)
            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)