File size: 6,768 Bytes
2a782f7
789add0
fb1c5b9
 
 
 
 
 
 
2a782f7
fb1c5b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
789add0
 
 
 
 
 
 
 
 
 
 
 
 
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
148
149
150
151
152
153
154
---
license: mit
task_categories:
- robotics
tags:
- robotics
- manipulation
- few-shot
- cross-embodiment
---

# AXIS held-out 20 — a cross-embodiment few-shot adaptation benchmark

20 tasks never seen in pretraining, 20 demonstrations each, LoRA adaptation, rollout evaluation.

**This release is the SPECIFICATION and the INDEX, not the demonstration data.** It is published
first and on purpose: everything here is what you need to *render* the benchmark on your own
embodiment, and none of it depends on our video encoding being finished.

> **Status: the task list is CANDIDATES.** The learnability gate (below) has not run yet. It can
> only remove tasks, never add them, so treat the 20 as an upper bound on the final set.

## What is being measured

Few-shot adaptation to a target setup — **not** generalisation. A policy gets 20 demonstrations of
an unseen task, adapts by LoRA, and is evaluated on that same task under **the same conditions the
demonstrations contain**.

That last clause decides a question that is otherwise answered wrongly. Measured from the
demonstration data itself:

| axis | varies across demos? | so at eval |
|---|---|---|
| object identity, placement, jitter | **yes** | vary |
| robot initial state | **no** — 0.22° / 1.8 mm across 856 and 1,148 episodes | **fix** |
| camera pose | **no** | **fix** |
| appearance / lighting / texture | **no** — the randomisation config holds position deltas only | **fix** |

Randomising something at eval that the demos never contain measures a regime the benchmark does
not train for, and produces a number that answers no one's question.

## Layout

```
spec/
  benchmark_spec.json    task list, operations, gate thresholds, protocol
  camera_rig.json        the REQUIRED wrist rig, and our measured one (they differ — see below)
  eval_seeds.json        the seed list every embodiment must use
franka/
  index/heldout20_index.jsonl    one row per demonstration
```

Embodiment-first, so adding an embodiment cannot disturb an existing one.

## The camera rig is PRESCRIPTIVE

A cross-embodiment benchmark must fix the rig, or a score difference between embodiments confounds
the robot with its cameras.

**Wrist camera — reproduce this, expressed relative to your arm's end-effector frame:**

```xml
<camera mode="fixed" name="eye_in_hand" pos="0.05 0 0" quat="0 0.707108 0.707108 0" fovy="75"/>
```

Taken verbatim from `robosuite/models/assets/robots/panda/robot.xml`, which is what LIBERO trains
and evaluates against.

**Our own Franka renders do NOT comply, and that is why the rig is specified rather than frozen.**
Derived from the render data by composing the inverse end-effector pose with the recorded
`camera_to_world`, and checked for rigidity (max elementwise std **7.3e-08** — the camera really is
bolted on, so the number is a measurement, not a fit):

| | required (LIBERO) | measured (our Franka) | delta |
|---|---|---|---|
| hand→camera translation | `[0.05, 0, 0]` m | `[-0.074, 0, 0.0292]` m | **12.7 cm, opposite side in x** |
| vertical FOV | **75°** | **51.87°** | **23.1° narrower** |

Both measured values have **zero spread across all 20 tasks**. Reproduce the required column.

**Third-person camera:** one fixed pose shared by every task and episode, `[0.95, 0.0, 1.43]`. Per
episode intrinsics and `camera_to_world` are in the index.

## The index

One row per demonstration. `split` is `adapt` (the 20 used for adaptation) or `spare`.

| field | meaning |
|---|---|
| `task_id`, `attempt_id` | **the join key.** Never join on `episode_index` — every build renumbers it |
| `source_episode_index` | which episode of our render this is |
| `n_frames`, `fps` | 15 fps |
| `cameras` | per camera `intrinsics` (3×3 K, row-major) and `camera_to_world` (4×4, row-major), on `adapt` rows |
| `camera_names` | all 10 rendered cameras: `frontview`/`wrist` plus four randomized replicas of each |
| `join_origin` | `published` (18 tasks) or `action_match` (2) — see below |
| `variant_id`, `seed`, `model_id` | scene identity |

`intrinsics` is K for the **source** resolution (640×360). If you render at another resolution,
adjust K rather than using it as-is.

### Two tasks have a recovered join

Tasks **815** and **1746** ship a `task_manifest.json` that parses with `records: []` while having
1,280 and 40 rendered episodes. Their attempt↔episode pairing was recovered by matching
trajectories directly. The method's accuracy was **measured, not assumed**: scored against task
868's published manifest it was **401/401 correct**. Rows carry `join_origin` so you can tell.

## Evaluation protocol

- **50 episodes per task**, using `spec/eval_seeds.json`. Every embodiment draws the same seeds, or
  a cross-embodiment comparison silently compares different layouts.
- Eval scenes vary object set / placement / jitter only.
- At 50 trials a true 20% rate carries a 95% CI of roughly ±11 points.

## The learnability gate (not yet run)

Quality filters say the *demonstrations* succeed. They say nothing about whether a policy can learn
the task from 20 of them — and the previous 10-task version of this benchmark floored, with **4 of
6 evaluated tasks at 0/10**. A benchmark whose tasks are all unlearnable ranks nothing.

Every task must satisfy, with `pi05_base` + 20-demo LoRA:

```
20% <= success rate <= 80%
```

Two-sided on purpose. Replacement depth is thin where it matters: `pick`, `place`, `reach` and
`remove` have exactly one eligible task each, so a gate failure there cannot be backfilled.

## Known limits, stated rather than discovered

- **Robot initial-state robustness is not measured.** The demos start from one fixed pose, so the
  benchmark cannot test it. This is our largest measured weakness elsewhere — on LIBERO-Plus, AXIS
  pretraining loses 30.9 points on Robot Initial States while gaining 4.5 on Camera Viewpoints and
  7.4 on Sensor Noise. That belongs to the pretraining corpus and is fixed by re-rendering *that*,
  not by perturbing this benchmark.
- **Visual robustness is not measured**, for the same reason: no appearance variation exists in the
  demonstrations.
- **`pick_place`, `stack`, `push`, `pull` are absent** — no task reaches the quality bar. A data
  limit, not an oversight.

---

## Also in this repository

This repo predates the held-out benchmark and carries other AXIS releases. They are unrelated to
the benchmark above and are not part of it:

- `stage1_5k_camera_fixed/` — stage-1 5k camera-fixed corpus and its artifacts (AWR weights,
  schedules, per-task manifests, roots/ranges, norm stats).
- the round-1 stage-2 LIBERO build (2,000 episodes / 338,575 frames / 128 px / LeRobot v3.0).

The held-out benchmark lives entirely under `spec/` and `<embodiment>/`.