File size: 7,360 Bytes
5610da4 ce7cc25 5610da4 ce7cc25 e3ca385 1d6d0ac e3ca385 1d6d0ac e3ca385 ce7cc25 5610da4 ce7cc25 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | # Render request — held-out benchmark, 20 tasks × 20 demos
Everything needed is in this folder. Read the attempt lists here, upload the rendered output here.
**What:** re-render 400 existing demonstrations (20 tasks × 20 attempts) for the **default Franka**,
at the **stage1_5k camera rig**, with **appearance pinned**.
### Why 20 and not 10 — and why the existing LIBERO-rig renders don't count
Ten of these tasks are currently in the stage1_5k pretraining data, so they cannot be scored against
today's models. They are listed as `status: conditional`. The other ten are `status: active`.
It would seem to follow that only the ten active tasks need rendering. It does not, for two reasons:
1. **The conditional ten become usable the moment a stage-1 run excludes the held-out set** — which
is planned. Rendering them now avoids a second round trip.
2. **Their existing renders don't meet the appearance requirement.** Tasks 810/815/929/... already
have renders at the correct camera rig in `Franka-Datasets-v2-5k-LIBERO`. We checked those
manifests: they run `TextureModder` and `LightingModder` too. So they need re-rendering for
appearance regardless of the rig.
So the ask is all 20. If you have to stage it, do the ten marked `active` first.
**Why re-render:** these demonstrations currently exist only at a superseded camera rig. The
pretraining corpus (`stage1_5k_camera_fixed`) uses a different one, and this benchmark evaluates
models pretrained on it — so a mismatch would give every evaluated model a train/eval viewpoint
gap rather than measuring the model.
---
## 1. The camera rig — must match exactly
Verified identical across sampled tasks of `Franka-Datasets-v2-5k-LIBERO` and
`Franka-Datasets-v2-30k-LIBERO`. Machine-readable copy: [`camera_rig.json`](camera_rig.json).
| | value |
|---|---|
| embodiment | **franka** (default) |
| camera group | **`camera_fixed`** |
| wrist camera | mounted on the **end-effector body**, hand→camera translation **`[0.05, 0.0, 0.0]`**, quat **`[0, 0.707108, 0.707108, 0]`**, **fovy 75°** |
| third-person camera | `camera_to_world` translation **`[1.0086, 0.0, 1.1904]`** |
| resolution | **640 × 360** |
| fps | **15** |
**The wrist rig is the one that matters most.** The superseded batch had it at
`[-0.074, 0, 0.0292]` with fovy 51.9° — 12.7 cm away, on the opposite side of the hand, and 23°
narrower. If a render comes back at the old rig it is unusable, and the difference is invisible in
a single still frame: it only shows up in how the image moves.
A quick self-check after rendering one episode: compose the inverse end-effector pose with the
recorded `camera_to_world` and confirm the translation is `[0.05, 0, 0]` to 3 decimals, and that it
is constant across frames (it should be rigid — ours measures a per-element std of ~7e-08).
---
## 1b. Appearance must be PINNED — this is new, and it matters more than the rig
The existing demos are **appearance-randomized per demo**: the render pipeline runs
`TextureModder` and `LightingModder` off a per-attempt `variant_seed`, so each of a task's 20
demonstrations has a different table material, a different floor and different lighting. We only
found this by looking at the frames — it is recorded in the render manifest, not in the task's
`domain_randomization` block, which contains position deltas only.
For a 20-demo benchmark that is a problem. The policy has to learn appearance-invariance *and* the
manipulation skill from 20 samples, so the score conflates two abilities and cannot be attributed
to the one under test. Measured on the current demos: a reach task scores 82% while two
manipulation tasks score 2% and 0%, with 99 of 100 rollouts running to the step cap — acting, never
completing.
**So please render all 20 demos of a task under ONE fixed appearance.** Everything in this list
held constant across every demo of every task:
| pin | why |
|---|---|
| **table material** | varies per demo today — six demos of task 953 have six different tables |
| **floor / background** | varies with it; one demo sits on brown parquet, another on green grass |
| **lighting** | `LightingModder` is active per demo |
| **object materials / textures** | `TextureModder` recolours the OBJECTS too, so the same mesh reads as a different object between demos |
Concretely: disable `TextureModder` and `LightingModder`, or hold `variant_seed` constant across
the whole batch.
**Object meshes do NOT need changing** — we verified those are already consistent. The eval scene
loads `cup_2`, `muffin_4`, `sweet_potato_4`, `tissue_box_2` from exactly the same
`tabletopgen/task_88200032_layout_1_rand_00000` directory the training scene references. The
problem is purely that their *materials* are randomized, so identical geometry looks like a
different object from demo to demo.
Object PLACEMENT should keep varying — that is the axis the benchmark is meant to test.
If your pipeline cannot pin any of these, please say so rather than varying them — we would rather
know and design around it than discover it in the frames again.
## 2. What to render
| file | contents |
|---|---|
| [`tasks.json`](tasks.json) | all 10 tasks, required + spare attempt ids, the rig |
| [`attempts/task_<id>.json`](attempts) | one file per task |
| [`attempts/all_attempts.csv`](attempts/all_attempts.csv) | flat table — `task_id, task_name, operation, role, attempt_id, n_frames, reward_final` |
See [`tasks.json`](tasks.json) and [`attempts/all_attempts.csv`](attempts/all_attempts.csv)
for all 20 tasks, each marked `active` or `conditional`.
**`attempt_id` is the identity — please render exactly those.** They are not arbitrary: they were
selected as the highest-reward demonstrations available for each task, and they are the only ones
with verified trace and goal data. Episode indices are renumbered by every build, so please key
your output on `attempt_id`, not on episode order.
**`role = spare`** lists 40 further attempts per task, ranked. If a required attempt fails to
render, take the highest-ranked spare and say which you substituted — don't silently drop one, as
20 demos is the whole adaptation budget and a missing demo is 5% of it.
---
## 3. Where to upload
Please upload into `uploads/` in this same folder, one directory per task:
```
uploads/
task_809/
meta/ LeRobot v3.0 metadata
data/ parquet
videos/ observation.images.third_person/ and .wrist/
render_manifest.json attempt_id -> episode_index, per episode
task_811/
...
```
`render_manifest.json` is the important one — without an `attempt_id → episode_index` mapping we
have to recover the pairing by matching trajectories, which works but is slower and needs checking.
If your pipeline already emits `task_manifest.json` in the usual shape, that is perfect; please just
make sure `records[]` is populated (two tasks in the previous batch shipped with `records: []` while
having over a thousand episodes).
---
## 4. Anything else
If a task can't be rendered at this rig for a structural reason, please say which and why rather
than substituting a different rig — we'd rather run a 9-task benchmark than one with a mixed rig.
Questions on any of this are welcome before rendering starts; a wrong rig is expensive to discover
afterwards.
|