EVALUATION.md

# Actual evaluation method and results

The historical benchmark is `robotwin-eval-subset-v2-seen-unseen`, fixed subset commit
`b5abe512dafd400736c6dca5905f2fd262dd9029`: 50 tasks × 10 clean/seen + 10 randomized/unseen
episodes, 1000 episodes per model. The same 1000 episode IDs and initial RGB/public-state
hashes match across endpoints. Accepted policy failures are included; infrastructure retries
are excluded. Evaluation uses unseen instructions, not unseen-instruction training.
All tasks are present in the all-clean training set. Seen/unseen refers to instruction pools;
clean/seen versus randomized/unseen changes both visual conditions and instruction pools.
These results concern the original fixed-table runs, not later low-table adaptations.

| Model | Clean / seen | Randomized / unseen | Overall |
|---|---:|---:|---:|
| ep2 | 381/500 (76.2%) | 204/500 (40.8%) | 585/1000 (58.5%) |
| ep6 | 416/500 (83.2%) | 178/500 (35.6%) | 594/1000 (59.4%) |

Per-task counts, completed/planned counts, success rates, worker counts and measured episode
wall times are in `RESULTS.json`. Wall times are specific to the actual distributed setup,
not hardware-independent model latency. These are historical results, not current service status.

Model inference used F32 weights/main path, eager action attention, compile=false, tf32=false,
autocast=false. Official FA2 vision casts FP32 Q/K/V to BF16 internally and restores output dtype.
E5 causal history is reset per episode/session, uses only confirmed executed command prefixes,
and prior public drive-target state observations. It is not measured-qpos history.

The checkpoint requires custom runtime loading, the original normalization, three camera inputs,
14-dimensional execution commands and baseline dependencies. Core runtime source is preserved
from original training archives. `scripts/e5_rollout_history.py` implements the causal history contract.
Use `HARDWARE.md` for actual inference/rendering machines and precision boundaries.

Exact original server, transport, renderer worker and subset verifier source from both evaluations is included and hashed in `SOURCE_MANIFEST.json`.

Exact end-to-end rollout reproduction still requires the fixed subset, compatible RoboTwin
simulation assets and matching server/renderer integration. No generic upstream evaluation
script is represented as the exact 1000-episode harness. No new rollout was run for this release.
原始记录保存在本地。

## Portable entry

Download the selected public checkpoint, including `training_config.yaml`, to `checkpoints/ep2`
or `checkpoints/ep6`. The release identity adapter verifies all six original shard hashes,
the index, sanitized model/training config and normalization before loading. It replaces
only checks tied to private training records. The loader now reads `training_config.yaml`
beside the shards. `SOURCE_MANIFEST.json` lists these changes and retains both original
registry hashes. The earlier training-archive rollout-history version is retained under
`training_reference/`; the active version is the exact source used by both evaluations.

Install the CUDA inference dependencies; keep Qwen3-VL under
`checkpoints/dependencies/Qwen3-VL-4B-Instruct`. Place compatible RoboTwin source/assets in
`third_party/RoboTwin`. Original warmup requires the clean adjust_bottle episode0 HDF5 under
`data/raw/adjust_bottle/aloha-agilex_clean_50/data/episode0.hdf5`.
The fixed subset must be a checksummed Git checkout. A public mirror is
[ar-mine/robotwin-eval-subset-v2-seen-unseen](https://huggingface.co/datasets/ar-mine/robotwin-eval-subset-v2-seen-unseen/tree/73277c2022be224d3afb195b503a3cff56ded46d),
at HF revision `73277c2022be224d3afb195b503a3cff56ded46d`. Its package SHA256SUMS is byte-identical
to the original benchmark (SHA256 `a8a41f7c5162732cce287e77628840bf754f631825d3a6a343d51c648f932eb9`).
The wrapper accepts either the original Git commit above or this fixed public mirror commit,
verifies the exact checksum-file hash and all package files, and records the actual checkout
commit used. Set `SUBSET_ROOT` to `benchmarks/robotwin_eval_subset_v2_seen_unseen` inside that
checkout and `RUN_DIR` to a fresh local result directory. Configure the
already-established transport so the renderer can reach the inference server on localhost
port 9340. The wrapper creates no SSH tunnels, scheduler allocations or remote jobs.

On the inference machine (A100), after selecting its GPU through `CUDA_VISIBLE_DEVICES`:

```bash
python scripts/release_eval.py server --epoch ep6 --output "$RUN_DIR/server" --port 9340 --run
```

On the renderer (RTX 5090), set `EVAL_GPU_UUID` and `CUDA_VISIBLE_DEVICES` to the same allocated
render GPU UUID. Execute an episode from the fixed subset:

```bash
python scripts/release_eval.py episode --epoch ep6 --subset "$SUBSET_ROOT"   --task adjust_bottle --setting clean --index 0 --output "$RUN_DIR/adjust_bottle_clean_0"   --port 9340 --run
```

Omit `--run` to validate/print a plan without starting a GPU process or service. For epoch2
choose `--epoch ep2` on both sides. The original worker applies per-episode seeds, RT rendering,
confirmed-action and public-state history, session reset and strict server-identity checks.
The wrapper runs one episode at a time; full1000 orchestration, resumability and infrastructure
retry accounting remain the executor's responsibility. Do not compare an incomplete run with
the historical full1000 table. Portable entries received CPU validation only; target-GPU
loading, the two-episode causal-history smoke, and full benchmark reproduction remain pending.