README.md

---
pretty_name: RobotAgentBench simulation assets
license: other
license_name: robotagentbench-asset-terms
license_link: LICENSE.md
viewer: false
tags:
- robotics
- embodied-ai
- simulation
- benchmark
---

# RobotAgentBench simulation assets

Version 0.1.7 supplies simulation assets, initial-state resources and the full
Context Reasoning question banks and standard answers for RobotAgentBench.
Evaluation code: https://github.com/AetherLabsAI/RobotAgentBench/tree/main.

The benchmark covers 18 short tasks, 8 long tasks, Task Discovery, Context
Reasoning, Failure Recovery, Robustness, and four OpenEnd household scenarios.
Isaac Sim and native SDK dependencies are installed separately.

## Shared asset release 0.1.1

Use this asset release with saved configuration bank **0.2.1**. It includes
the shared OpenEnd room layout, drawer-reference layer, and resource bindings.
The 18 original archive parts are unchanged; two additional archives total
approximately **27.5 MB**. The complete download is approximately **4.57 GB**
across 20 archives. Update the evaluator and reuse verified downloads when
upgrading, but install into fresh asset and
runtime directories. Earlier asset and configuration releases remain available.

## Shared asset release 0.1.2

Use this release with the current evaluator and saved configuration bank 0.2.1.
It adds the original shelf scene-preparation dependencies and complete fridge
interior resources. The two supplements total approximately 7.3 MB; all earlier
archives are reused unchanged. Python helper sources and their notices ship in
the code repository; native third-party payloads retain their original licenses.
Use fresh asset/runtime directories and reuse verified cached downloads.

## Shared asset release 0.1.3

This release includes the original scene-preparation metadata for waste-selection
and table-tidying tasks. The supplement is approximately 438 KB; all 22 earlier
archives are reused unchanged. The complete download is approximately **4.58 GB**
across **23 archives**, including the license archive. Use fresh asset/runtime
directories and reuse verified cached downloads.

## Shared asset release 0.1.4

This release completes shared room resources for saved Failure Recovery configurations.
The room bases are stored once; configuration records contain lossless scene
patches that reconstruct the saved room state. All 23 earlier archive parts
are reused unchanged. The complete asset download is approximately **6.66 GB**
across **30 archives**. Use fresh asset and runtime directories and reuse
verified cached downloads.

## Shared asset release 0.1.5

This release includes all **50 original room layouts** and their runtime
dependencies. Layouts, meshes, textures, and shared resources are stored
once and reused across saved configurations. Original file hashes are
preserved; scoring, physics, and saved initial states are unchanged.
The complete download is approximately **7.04 GB**
across **31 archives**. All 30 earlier archives are
reused unchanged. Update the evaluator, reuse verified cached downloads,
and install into fresh asset and runtime directories. Use configuration
bank **0.2.5** with this release.

## Shared asset release 0.1.6

Use this release with saved configuration bank **0.2.6**. It adds one compact,
lossless shared scene-reconstruction dependency for the new Sofa recovery
configuration. The supplement is approximately **515 KB**; all 31 earlier
archives, all 50 room layouts, and third-party notices remain unchanged.
The complete asset download is approximately **7.04 GB** across **32 archives**.
Reuse verified downloads and install into fresh asset/runtime directories.

## Shared asset release 0.1.7

Use this release with saved configuration bank **0.2.17**. It adds one compact,
lossless reconstruction dependency for the new Sofa recovery configuration.
The supplement is approximately **516 KB**; all 32 earlier archive parts,
all 50 room layouts, and third-party notices remain unchanged. The complete
download remains approximately **7.04 GB**, across **33 archives**. Reuse
verified downloads and use fresh asset/runtime directories.

## Install

Install the evaluator from its code repository, then use its versioned manifest:

```sh
robotagentbench assets fetch resources/assets-0.1.7.json --cache downloads
robotagentbench assets install resources/assets-0.1.7.json --cache downloads --destination Assets
robotagentbench assets configure resources/assets-0.1.7.json --destination Assets --runtime Runtime --profile room
```

For offline installation, copy all release archive parts into `downloads`.
Do not concatenate them. The installer verifies archive and member hashes.
Use separate fresh runtime directories for `room`, `atomic_coffee`,
`failure_recovery`, and `cloth` profiles. Standard answers are evaluator data;
they are not part of the agent observation interface.

## Saved initial configurations

The separate configuration release contains **1,470 saved initial configurations**:
30 for each of 49 task or scenario groups. These are release configurations for
reproducible evaluation, not the exact instances used for the paper's reported
results. Each record retains its saved seed and initial state.

Configuration release **0.2.17** contains 30 saved configurations for each
of the four OpenEnd scenarios. Household memory and the high-level request
are accompanied by **Today's situation**, covering weather, trip timing,
confirmed plans, and personal preferences.

Stains, empty containers, used packaging, object locations, and cleanliness
must be discovered from camera observations, not disclosed in the daily context.
Variants that differ only in visible state use the same public text. The six
context/scene variants per scenario each have five spatial initial states;
they are not thirty distinct stories.

Version 0.2.6 updates three saved initial states using original-rule seeded
replacements that passed native initial admission: lighting episode 4,
combined coffee disturbance episode 5, and sofa recovery episode 2. The
other 1,467 states, public contexts, task definitions, scoring rules, and
physical parameters are unchanged from 0.2.5; its two earlier replacements
are retained. Earlier versions remain available.
Version 0.2.7 updates two further saved initial states: coffee-bag disturbance
episode 6 and combined coffee disturbance episode 6. Both replacements passed
their original native initial admission. The other 1,468 states are unchanged
from 0.2.6; public contexts, task definitions, scoring, and physical parameters
are unchanged. Shared assets remain at version 0.1.6.

Version 0.2.8 updates three further saved initial states: combined coffee
disturbance episode 7, button-lighting episode 8, and sofa recovery episode 3.
Each replacement passed its original native initial admission. The other
1,467 states are unchanged from 0.2.7; public contexts, task definitions,
scoring, and physical parameters are unchanged. It uses shared assets 0.1.7.

Version 0.2.9 updates nine further saved initial states: table-waste clearing
episode 23; microwave heating and retrieval episodes 14, 17, and 28;
dinner preparation episode 3; and long tasks 17/26, 19/21, 21/19, and 22/25
(task/episode). Each replacement passed its original native initial admission.
The other 1,461 states are unchanged from 0.2.8; public contexts, task definitions,
scoring, and physical parameters are unchanged. Shared assets remain at 0.1.7.

Version 0.2.10 updates three further saved initial states: preparing to leave
(OpenEnd) episode 2, and button-lighting episodes 13 and 18. Each replacement
passed its original native initial admission. The other 1,467 states are
unchanged from 0.2.9; public contexts, task definitions, scoring, and physical
parameters are unchanged. Shared assets remain at 0.1.7.

Version 0.2.11 updates eight further saved initial states: four Task Discovery
instances and four OpenEnd instances. Each replacement passed its original
native initial admission. The other 1,462 states are unchanged from 0.2.10;
public contexts, task definitions, scoring, and physical parameters are
unchanged. Shared assets remain at 0.1.7.

Version 0.2.12 updates four further saved initial states: two Task Discovery
instances, one OpenEnd tidying instance, and one coffee-machine displacement
instance. Each replacement passed its original native initial admission.
The other 1,466 states are unchanged from 0.2.11; public contexts, task
definitions, scoring, and physical parameters are unchanged. Shared assets
remain at 0.1.7.

Version 0.2.13 updates two further saved initial states: coffee-machine
displacement episode 4 and button-lighting episode 21. Each replacement
passed its original native initial admission. The other 1,468 states are
unchanged from 0.2.12; public contexts, task definitions, scoring, and physical
parameters are unchanged. Shared assets remain at 0.1.7.

Version 0.2.15 updates three further saved initial states using the original
sampling and admission rules: combined coffee disturbance (episode 21, seed 204),
coffee-button lighting (episode 14, seed 73), and sofa Failure Recovery
(episode 19, seed 94). The other 1,467 saved states and shared assets 0.1.7 are
unchanged. Use RobotAgentBench source commit
`f86a908d709fb94dcfc970be8c5da93305be9968` or later, which supports the nested
saved-scene dependencies in this bank. These public configurations differ from
the paper's evaluation instances.

Version 0.2.14 updates fourteen further saved initial states: ten coffee
interference variants and four OpenEnd episode configurations. Each replacement passed
its original native initial admission. The other 1,456 states are unchanged
from 0.2.13; public contexts, task definitions, scoring, and physical parameters
are unchanged. Shared assets remain at 0.1.7.

Version 0.2.16 updates two further sofa Failure Recovery initial states:
episode 8 (seed 96) and episode 30 (seed 110). Both replacements passed their
original native initial admission and the evaluator's saved-state restoration.
The other 1,468 saved states, public contexts, task definitions, scoring rules,
physical parameters, and shared assets 0.1.7 are unchanged. Use RobotAgentBench
source commit `f86a908d709fb94dcfc970be8c5da93305be9968` or later. These
public configurations differ from the paper's evaluation instances.

Version 0.2.17 updates six further saved initial states using the original
sampling and admission rules: microwave heating and retrieval (episode 11,
seed 34), Task Discovery dinner (episode 17, seed 131), sofa Failure Recovery
(episode 6, seed 83), coffee preparation (episode 27, seed 39), bread preparation
(episode 6, seed 31), and meal-tray carrying (episode 6, seed 31). All six
replacements passed task-owned native initial admission and saved-state
restoration/readonly verification. The other 1,464 state records and shared
assets 0.1.7 are unchanged. All 1,470 saved initial states have passed their
original task-owned admission checks; this does not claim completed robot
tasks or model success. Use evaluator commit
`e8448f732572cacf330554c3620d6f25341be81f` or later for lossless external
saved-scene attributes. Historical releases remain available.

Task-owned attachments, original spawn receipts, and lossless room patches
reconstruct the saved configurations. Room meshes and textures remain in
the shared asset download.
The configuration download is approximately **260 MB**, split into seven archives.

Earlier releases remain available under `saved-instances/0.1.0/` and
`saved-instances/0.2.0/`, `saved-instances/0.2.1/`, `saved-instances/0.2.2/`, and
`saved-instances/0.2.3/`, `saved-instances/0.2.4/`, `saved-instances/0.2.5/`, `saved-instances/0.2.6/`, and `saved-instances/0.2.7/`, and `saved-instances/0.2.8/`, and `saved-instances/0.2.9/`, and `saved-instances/0.2.10/`, and `saved-instances/0.2.11/`, and `saved-instances/0.2.12/`, and `saved-instances/0.2.13/`, and `saved-instances/0.2.14/`, and `saved-instances/0.2.15/`, and `saved-instances/0.2.16/`. Install 0.2.17 into a new directory when upgrading.

Install the shared assets once. Configurations store task parameters, poses,
robot initialization, required cloth-state arrays, and compact lossless scene
patches; they do not duplicate complete room geometry and textures for every episode.

```sh
python -m robotagentbench.simulation.saved_instance_release.download \
  --manifest resources/saved-instances-0.2.17.json \
  --cache downloads/configurations --destination /absolute/path/to/SavedConfigurations
```

See the [configuration guide](https://github.com/AetherLabsAI/RobotAgentBench/blob/main/docs/INSTANCES.md)
for selection and restoration, and the repository's InternVLA and Codex examples
for connecting an agent. The fine-tuned InternVLA model is available at
[AetherLabs-AI/InternVLA-A1.5-RAB](https://huggingface.co/AetherLabs-AI/InternVLA-A1.5-RAB).

## Licenses

Team-authored materials use CC BY 4.0. Third-party resources retain their own
terms, including CC BY-NC-SA 4.0 for the AgiBot GenieSim robot resources.
See [LICENSE.md](LICENSE.md), [NOTICE.md](NOTICE.md), and `LICENSES/`.
The license archive is included in the installer manifest so these notices
accompany automatic asset downloads. The code's Apache-2.0 license does not
relicense simulation assets or third-party materials.