Add planner-only group-relative LoRA post-training

[hova88]

Summary

• add an opt-in PlanningHeadGRPO with LoRA adapters and Gaussian displacement rollouts
• freeze stage-2 UniAD when planner_posttrain_only=True; the default PlanningHeadSingleMode path is unchanged
• use detached-action group-relative policy gradients and exact non-negative reference KL
• include one config, focused tests, and scoped nuScenes v1.0-mini evidence

Validation

• 9 passed in tests/test_planner_grpo.py
• synthetic planner forward/backward gives non-zero LoRA-B and log_std gradients
• official uniad_base_e2e.pth planner weights load with no missing wrapped base weights
• MMCV config parsing passes

Preliminary evidence

On 12 nuScenes v1.0-mini open-loop frames, the earlier LoRA/auxiliary-fit campaign reached 3.60 m mean / 0.98 m median deterministic mean-path ADE. Best logged rollout minFDE was 0.052 m. The earlier campaign's PPO term was inactive; these values are therefore not presented as a GRPO gain. No trainval or closed-loop claim is made.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fb1e688dce

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[hova88]

detail experiment link: https://hova88.github.io/nano.UniAD/planner_grpo_lora/