christina040807/rlds_cw_v1

OdorSim OVLA Physical Two-Cup RLDS v1

This repository contains an odor-aware RLDS/TFDS dataset generated with OdorSim, robosuite/MuJoCo, GADEN, and a simulated continuous eight-channel MOX e-nose. It supports training an odor-conditioned OpenVLA/OVLA policy for a two-cup sniff, reject-or-grasp, and place task.

Dataset

TFDS name:       odor_ovla_physical_full_v3_stratified
version:         1.1.0
episodes:        1,500
train:           1,350 episodes / 289,741 retained steps
validation:      150 episodes / 32,282 retained steps
ordered pairs:   all 30 unequal ordered gas pairs

Target gases and target-first/target-second order are balanced in both splits. Splits are made by complete episode.

Gas order:

ethanol, methanol, acetone, acetic_acid, ethyl_acetate, toluene

Sensor order:

101, 203, 302, 401, TGS2600, TGS2602, TGS2610, TGS2611

Observation Contract

Each step contains:

observation.image               uint8[256,256,3]
observation.wrist_image         uint8[256,256,3]
observation.state               float32[8]
observation.odor                float32[100,8]
observation.odor_padding_mask   bool[100]
observation.odor_valid          bool
action                          float32[7]
language_instruction            string

observation.odor is a causal five-second MOX history sampled at 20 Hz. The mask follows Transformer key-padding semantics:

true  = left-padding; ignore the timestep
false = real MOX sample

The RLDS stores raw MOX histories rather than precomputed ScentFormer embeddings. Run the frozen odor encoder on odor using odor_padding_mask, then provide only the latent odor embedding to OVLA.

Sparse sniff HOLD landmarks are retained throughout the sniff period. A downstream loader must not remove all zero-delta actions, because doing so would delete the supervision that teaches OVLA to wait while smelling.

Loading

After downloading the repository snapshot:

import tensorflow_datasets as tfds

builder = tfds.builder_from_directory(
    builder_dir=(
        "odor_ovla_physical_full_v3_stratified/1.1.0"
    )
)
train = builder.as_dataset(split="train")
validation = builder.as_dataset(split="val")

Integrity

The TFDS version directory includes SHA256SUMS. From that directory, run:

sha256sum -c SHA256SUMS

All 146 listed files should report OK.

See HANDOFF.md for detailed physical validation, schema notes, and known limitations. The generating code is on the cw_sensor_simulation branch of the OdorSim repository.