LACE: Language-conditioned Avoidance in a Chemistry-lab Environment

LACE is a real-robot manipulation dataset collected on a single-arm SO-100 in a chemistry-lab setting. It targets language-conditioned manipulation where the instruction — not vision alone — determines the correct behavior. Its defining feature is mirror-pair avoidance: two tasks share an identical scene, and only the wording says which object to move and which to avoid (e.g. "transfer the beaker, avoid the cylinder" vs. "transfer the cylinder, avoid the beaker"), making language strictly load-bearing.

At a glance

• Robot: SO-100 follower arm - 30 fps
• Cameras (3): top (OBS_IMAGE_1), side (OBS_IMAGE_2), wrist (OBS_IMAGE_3) - 640x480
• Tasks: 17 labware-manipulation tasks
• Demonstrations: 1,716 human teleoperation episodes - 643,773 frames
• Format: LeRobot v2.1 (one self-contained dataset per task, under train/)
• Language: every task ships a base instruction, 4 in-distribution paraphrases, and 1 verb-disjoint held-out rewording (see paraphrases.json)

Task groups

Tasks span five groups; the obstacle-avoidance group (star) is the core of the benchmark -- there, language alone disambiguates the target from the distractor.

Group	#	Description
A	3	Atomic pick / grab / place
B	3	Source-target visual disambiguation (size / color / content distractor)
C	3	Verb diversity (pour / open / place)
D (star)	7	Obstacle avoidance (transfer X, avoid Y), including the mirror pair
E	1	Special manipulator (clamp / magnetic)

Tasks

#	Task	Group	Episodes	Frames
1	pick_beaker	A	100	38,929
2	grab_centrifuge_tube	A	104	27,920
3	pick_centrifuge_tube_to_yellow_rack	A	101	50,630
4	pick_small_beaker_into_large	B	104	36,989
5	pick_tubes_transparent_to_yellow_rack	B	100	39,981
6	pick_color_reagent_tube_to_yellow	B	100	51,404
7	pour_liquid_small_to_large	C	104	57,859
8	open_petri_dish_lid	C	102	32,935
9	place_graduated_cylinder	C	100	32,292
10	transfer_tubes_avoid_beakers	D	101	30,472
11	transfer_cylinder_avoid_other_cylinders	D	99	25,202
12	grab_color_tubes_avoid_measuring_cylinders	D	100	32,498
13	transfer_cylinder_avoid_large_beaker	D	100	23,899
14	place_tube_avoid_beaker	D	101	27,720
15	transfer_beaker_avoid_cylinder	D	99	25,117
16	move_funnel_avoid_beaker	D	101	27,726
17	clamp_magnetic_stir_bar	E	100	82,200

Layout

LACE/
├── paraphrases.json                 # all 17 tasks: base + 4 paraphrases + 1 held-out
└── train/
    └── <task_name>/                 # a self-contained LeRobot v2.1 dataset
        ├── data/                    # per-episode observation/action parquet
        ├── videos/                  # 3 camera streams (OBS_IMAGE_1/2/3)
        └── meta/                    # info.json, tasks.parquet, stats.json, episodes/, paraphrases.json

Each train/<task> is an independent LeRobot v2.1 dataset. Download the repo and load any task locally:

from huggingface_hub import snapshot_download
local = snapshot_download("littlewhite123/LACE", repo_type="dataset")
# each `train/<task>` folder is a self-contained LeRobot v2.1 dataset

Paraphrase protocol

Policies are trained on the base wording and evaluated on the held-out rewording -- which shares no main verb with any training wording -- to measure paraphrase robustness. The four in-distribution paraphrases support augmentation and test-time canonicalization studies. All instruction variants live in paraphrases.json, keyed by task name, e.g.:

{
  "transfer_beaker_avoid_cylinder": {
    "base": "Transfer the beaker to the designated position without touching the graduated cylinder during the process.",
    "paraphrases": ["Move the beaker to the designated position while keeping clear of the graduated cylinder.", "..."],
    "held_out": "Take the beaker to the designated position without contacting the graduated cylinder."
  }
}

License

Released under CC BY 4.0.

Citation

A citation will be added upon publication of the associated paper.