当前状态：研究归档。 StartShift-VLA 研究视觉—语言—动作（VLA）策略是否依赖训练演示中常见的默认初始状态。公开仓库保留核心代码、历史仿真基线与研究复盘；RISE-E、RISE-EA 和 RISE-EAR 仍是未验证的原型，其后续开发已经停止。项目说明
关键决策：来源核验通过后才能训练
训练尝试停在数据来源核验：已有 episode 元数据不足以确认其对应的 RobotInit 变体。保存的 G2 记录明确写明 gpu_training_started=false，因此没有完成基线微调与 RISE 的匹配数据比较。阻塞记录
这份归档的价值是让已观察的现象、未完成的假设和证据缺口可以被分别检查。进一步研究需要配对评估、可追溯的初始状态、多个模型与失败分类；这些是复盘中识别的缺口，并非本项目已经完成的结果。研究复盘
研究问题与保留证据
实验使用同一个 SmolVLA 基础检查点，在 vanilla LIBERO 的任务子集和 LIBERO-Plus 的 Robot Initial States 子集上记录策略表现。公开的主要评估共 920 条保存记录：ID 组 120 条、RobotInit 组 800 条。每组任务或变体运行 10 个 episode。实验条件
| 历史仿真评估 | 任务 / 变体组 | 成功次数 / 总次数 | 实测成功率 |
|---|---|---|---|
| ID：vanilla LIBERO 子集 | 12 | 86 / 120 | 71.67% |
| RobotInit：LIBERO-Plus 审计子集 | 80 | 36 / 800 | 4.50% |
封面图根据公开汇总数据重绘。两组成功率的描述性差异为 67.17 个百分点。两组没有按基础任务和物理场景配对，因此不能把全部差异解释为初始状态变化的因果效应。 结果也仅覆盖一个策略与一套记录的种子安排。
能说明什么，仍缺少什么
这些历史结果与初始化敏感性的假设一致，使重置条件下的可靠性成为值得继续研究的问题。公开记录同时保留了限制：完整限制说明
- 920 条主要记录均未保存初始状态向量和视频引用，无法仅凭任务标签重建物理状态配对。
- 798 次失败仍为
UNLABELED，尚无完整的失败机制分类。 - 未完成第二个模型、多种子复验或真机验证，不能由此概括所有 VLA 策略。
- 未验证 RISE 的方法收益；方法效果记为 N/A，不能解释为有效或无效。
如何复核
公开仓库提供 python scripts/verify_archive.py，用于检查保存文件的哈希、920 条记录、成功次数、分组与种子安排，并从已有记录重算指标。它复核保存证据，不重新运行仿真器或训练。 本页引用固定提交 8fb03bb；没有据此声称完成新的机器人实验、论文或部署系统。复核说明
Status: research archive. StartShift-VLA asks whether a vision-language-action (VLA) policy depends on the default reset states common in its training demonstrations. The public repository preserves core code, historical simulation baselines, and a research reflection. RISE-E, RISE-EA, and RISE-EAR remain unvalidated prototypes; their further development has stopped. Project overview
Key decision: require provenance before training
The training attempt stopped at a provenance check: available episode metadata could not certify the requested RobotInit variant membership. The retained G2 receipt states gpu_training_started=false. A matched-data comparison of baseline fine-tuning and RISE was therefore not completed. Blocked-run receipt
The archive makes the observed behavior, unfinished hypotheses, and evidence gaps independently inspectable. Stronger follow-up research would require paired evaluations, traceable initial states, multiple models, and failure classification. These are identified gaps rather than completed project results. Research reflection
Question and retained evidence
The evaluations used the same base SmolVLA checkpoint on a vanilla LIBERO task subset and the Robot Initial States subset of LIBERO-Plus. The principal public evidence contains 920 saved evaluation records: 120 ID episodes and 800 RobotInit episodes, with ten trials per task or variant group. Experimental conditions
| Historical simulation evaluation | Task / variant groups | Successes / episodes | Observed success rate |
|---|---|---|---|
| ID: vanilla LIBERO subset | 12 | 86 / 120 | 71.67% |
| RobotInit: LIBERO-Plus audit subset | 80 | 36 / 800 | 4.50% |
The cover chart is redrawn from the public summary data. The descriptive difference is 67.17 percentage points. The selections were not paired by base task and physical scene, so the entire difference cannot be attributed causally to reset state. The evidence covers one policy and one recorded seed schedule.
Interpretation and remaining gaps
The historical observations are consistent with initialization sensitivity and make reset robustness a concrete reliability question. The archive also preserves its limitations. Full limitations
- All 920 principal rows lack saved initial-state vectors and video references. Task labels alone cannot reconstruct physical-state pairing.
- All 798 failed episodes remain
UNLABELED; a complete failure-mechanism taxonomy is missing. - No second model, independent multi-seed replication, or physical-robot validation was completed. The result does not generalize to all VLA policies.
- RISE benefits were not validated. Its method effect is N/A, which establishes neither effectiveness nor ineffectiveness.
Verification boundary
The public entry point, python scripts/verify_archive.py, checks saved-file hashes, all 920 records, success counts, group identities, and the recorded seed schedule, then recomputes metrics from existing records. It verifies saved evidence; it does not rerun simulation or training. This page references commit 8fb03bb and claims no new robotics experiment, completed paper, or deployed system. Verification notes