reactor-x2-GR1-Manipulation-Task-v3

31 episodes, 2,656 frames of GR1 humanoid data, each episode built by adding a different food item onto the plate/turntable inside the microwave in one real recorded nvidia/Arena-GR1-Manipulation-Task-v3 episode, using Reactor XMAX X2 video editing. Standard LeRobot v2.1 layout (meta/, data/, videos/ at the repo root), so it loads directly with:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("kabilanKB/reactor-x2-GR1-Manipulation-Task-v3")

Source data

nvidia/Arena-GR1-Manipulation-Task-v3 — GR1 humanoid, IsaacLab-Arena, task "Reach out to the microwave and open it." 50Hz, 512x512 ego-view, teleop + MimicGen demonstrations, CC-BY-4.0. The source scene contains exactly two objects — the microwave and the robot — and no food: the pale disc visible inside is an empty turntable.

One demo episode (episodes/gr1_ep0, 94 frames, GR1 26-DoF action/state) was converted from the source LeRobot recording and used as the common base for every scene below.

What we did

Since the source has no food in it, every episode here re-renders the same 94-frame reach-and-open trajectory through Reactor X2 with a prompt that names one food item sitting on the plate/turntable and nothing else — "<food> on the white plate inside the microwave, warm interior glow, everything else unchanged". That "food only, minimal styling" phrasing is the one template (of five tried) that survives on this footage: naming a lighting/atmosphere style at the same time as new food competes for the same canvas and erodes structure coverage below use (see this project's PROJECT.md, "Adding also competes with restyling").

The task, instruction, and recorded action/state trajectory are unchanged in every episode — the robot only ever reaches for and opens the microwave door, it never touches the food. The food is scene dressing the policy should learn to ignore, not a new target.

Each edited batch is gated on structure_coverage (are the source edges still present in the edit?) with a floor of 0.75; a scene that drops below the floor is dropped entirely rather than shipped with corrupted labels. Of 32 food items attempted, 31 survived the gate — dhokla scored 0.748 and was dropped. Frames X2 does not return (~8% typical) are dropped from that episode along with their actions rather than reconstructed.

Food items (31 kept)

Poha, Oatmeal, Rice, Khichdi, Upma, Baked Potatoes, Sweet Potatoes, Steamed Broccoli, Steamed Carrots, Steamed Cauliflower, Corn on the Cob, Frozen Peas, Frozen Mixed Vegetables, Papad, Macaroni and Cheese, Omelette, Scrambled Eggs, Poached Eggs, Cheese Quesadilla, Ready-to-Eat Curry, Instant Noodles, Cup Pasta, Instant Soup, Microwave Popcorn, Veggie Nuggets, French Fries, Samosas, Spring Rolls, Mug Cake, Brownies, Carrot Halwa.

(Dhokla was attempted and dropped — see Caveats.)

Dataset stats

Episodes	31 (of 32 food items attempted)
Frames	2,656
FPS	50
Action / state dims	26 (GR1 arms + hands)
Camera	observation.images.ego_view, 512x512
Mean per-frame coverage	0.857 (min 0.765, max 0.926)
Mean residual drift	0.92px
Frames kept vs. pushed	2,656 / 2,914 (91%)

Per-episode provenance (scene id, prompt, coverage, drift, source frame range) is recorded in meta/augmentations.jsonl.

Caveats

• This augments an existing action-labeled episode; X2 is pixels-only and contributes no actions, proprioception, or camera pose of its own.
• X2 does not echo back the per-frame tag used to re-pair edited frames with their source (user_data), so frames are re-paired by arrival order under keep_backlog=true. This project's own validation run on this exact episode (out_gr1_food/report.json) rates this CONDITIONAL, not a full GO, for that reason — arrival-order pairing is likely correct but not guaranteed frame-for-frame.
• structure_coverage and drift are automated proxies for "does the label still match the pixels," not a guarantee of visual quality or that the intended food is what actually got painted in every frame.
• Whether training on this augmented data actually improves a policy is unproven — no A/B run against the un-augmented source has been done.
• Built from a single source episode (all 31 output episodes share one underlying trajectory); it adds visual diversity, not trajectory diversity.