| --- |
| license: mit |
| library_name: pytorch |
| tags: |
| - robotics |
| - vision-language-action |
| - florence-2 |
| - qwen3-vl |
| - bridge |
| - importance-sampling |
| --- |
| |
| # ReTime / OPTR — four-setting handoff |
|
|
| This package contains the server-authoritative Florence–Bridge OPTR code, its complete 20K common-start state and Uniform/Residual-OPTR 20.5K evaluation snapshots. It also contains the complete Qwen3-VL-4B–Bridge 29K common-start state and the environment/seed contract needed for paired multi-seed continuation experiments. |
|
|
| ## Unified checkpoint layout |
|
|
| The release is being normalized under [`settings/`](settings/README.md). Each of the four settings has the same entries: `common_start`, `uniform`, `residual`, `keyframe`, `frameskip`, `loss_is`, `configs`, and `dependencies`. |
|
|
| - [`settings/florence_bridge/`](settings/florence_bridge/README.md) |
| - [`settings/florence_robotwin/`](settings/florence_robotwin/README.md) |
| - [`settings/qwen_bridge/`](settings/qwen_bridge/README.md) |
| - [`settings/qwen_robotwin/`](settings/qwen_robotwin/README.md) |
|
|
| The common-start entry is the resumable branch point. Method endpoints are evaluation-ready model snapshots. Shared source code is kept once at repository root and referenced by each setting to avoid four inconsistent copies. |
|
|
| ## Rebuttal 训练与测评种子 |
|
|
| | 设置 | Common-start 训练 seed | Continuation 训练 seed | Evaluation policy seeds | |
| |---|---|---|---| |
| | Florence–Bridge 20K → 20.5K | `42` | 主实验 `117465`，已报告三组 `117465, 533389, 944964` | `518957, 765353, 686765, 516408` | |
| | Qwen3-VL-4B–Bridge 29K 与 29.5K | `42` | 历史 +500 配置为 `42`，继承 29K RNG。新多种子清单待定 | `518957, 765353, 686765, 516408` | |
|
|
| 完整的 Florence RNG 字段、Rebuttal 复用规则和原始记录链接见 [Bridge 训练与测评种子总表](BRIDGE_SEED_PROTOCOL.md)。 |
|
|
| ## What is included |
|
|
| - [`real_robot/`](real_robot/README.md)：Xtrainer2 真机数据转换、训练、双臂动作适配和 policy server 源码，以及中文交接说明。 |
| - `common_start_20k/checkpoints/steps_20000_accelerate/`: **full resumable state** (model, optimizer, scheduler metadata and per-rank RNG state). Start every new continuation arm here. |
| - `checkpoints/uniform/steps_20500_pytorch_model.pt`: model-only Uniform +500 snapshot. |
| - `checkpoints/residual/steps_20500_pytorch_model.pt`: model-only Residual-OPTR +500 snapshot. |
| - `proposals/residual_temporal_proposal_k8_alpha0.2.json`: frozen K=8, alpha=0.2 Residual proposal. |
| - `manifests/`: Bridge/RT1 provenance, content fingerprints and valid-anchor identity. |
| - `configs/`, `scripts/`, `starVLA/`: portable training and evaluation code. |
| - `EXPERIMENT_CONTRACT.md`: exact experimental contract and results. |
| - `AGENTS.md`: fast handoff instructions for an AI coding agent. |
| - `qwen_bridge_common_start_29k/checkpoints/steps_29000_pytorch_model.pt`: Qwen3-VL-4B–Bridge 29K model parameters. |
| - `qwen_bridge_common_start_29k/checkpoints/steps_29000_training_state.pt`: complete optimizer, scheduler, Python/NumPy/PyTorch/CUDA RNG and data-order state. |
| - `qwen_bridge_common_start_29k/config.full.yaml`: server-authoritative configuration that produced the 29K common start. |
| - `qwen_bridge_common_start_29k/ENVIRONMENT.md`: exact Qwen software and training environment. |
| - `qwen_bridge_common_start_29k/MULTISEED_PROTOCOL.md`: strict separation of common-start, continuation-training and evaluation seeds. |
| - `qwen_bridge_common_start_29k/AGENTS.md`: Qwen handoff constraints for an AI coding agent. |
|
|
| The 20.5K files are sufficient for evaluation but **not** for exact optimizer/RNG continuation. Any new branch must resume from the complete 20K state. |
|
|
| For Qwen–Bridge, exact continuation requires **both** 29K `.pt` files. |
|
|
| This release does not use implementation, code-path, absolute-path, current-manifest or self-referential SHA gates. Experimental identity is enforced through loadability and semantic model/data/episode/anchor/support/batch/budget checks. |
|
|
| ## Florence–Bridge headline result |
|
|
| Fixed no-video SimplerEnv WidowX/Bridge panel: four policy seeds (`518957`, `765353`, `686765`, `516408`) × four tasks × 24 episodes = 384 rollouts. |
|
|
| | Method | Stack | Carrot | Spoon | Eggplant | Overall | |
| |---|---:|---:|---:|---:|---:| |
| | Uniform +500 | 13/96 | 31/96 | 50/96 | 68/96 | 162/384 = 42.19% | |
| | Residual-OPTR +500 | 24/96 | 45/96 | 83/96 | 76/96 | 228/384 = 59.38% | |
|
|
| ## Florence–Bridge quick start |
|
|
| 1. Download this repository and install `requirements.txt` in the same software environment used for StarVLA. |
| 2. Download Florence-2-large and the OXE LeRobot Bridge/RT1 data. |
| 3. Point the environment variables listed in `AGENTS.md` to those local assets. |
| 4. Keep the contract fixed: two GPUs, BF16, ZeRO-2, micro-batch 4/GPU, accumulation 8, global batch 64, continuation seed 117465. |
| 5. Run one arm: |
|
|
| ```bash |
| bash scripts/run_florence_bridge_arm.sh uniform uniform-reproduction |
| bash scripts/run_florence_bridge_arm.sh residual residual-reproduction \ |
| proposals/residual_temporal_proposal_k8_alpha0.2.json |
| ``` |
|
|
| The launch path performs only file-presence and semantic data/support checks. |
|
|
| ## Important implementation facts |
|
|
| - `framework.name: QwenGR00T` is a historical registry name; the base VLM is Florence-2-large. |
| - The actual Florence–Bridge run used `cross_attention_dim: 1024` and DeepSpeed ZeRO-2. |
| - The optimizer groups are the VLM interface and action model. No explicit freeze list was configured; the executed action path updated the Florence vision/text interface and action head, while decoder parameters outside that path did not receive updates. |
| - Residual sampling uses ordinary `p/q` objective correction. This release intentionally does not upload Matched-Random/shuffle assets. |
|
|
| See `SOURCE_INVENTORY.md` before changing any method or training field. |
|
|
