Sim-Real-AAH/gah-sim-videos

A²H: simulation rollouts (before / after adversarial design)

Videos from the simulation training stage of Learning to Anticipate Physical Failures via Agentic Adversarial Harness (A²H), CoRL 2026 Late-Breaking Results. Project page: https://a2h-robot.github.io

During training, a coding agent writes a program that solves a task, then designs a variation of the scene that breaks that program (an augmentation), repairs the program, and writes down what it missed. These clips show each program on the original task and on every variation, before and after the repair. One seed per clip.

Layout: sim/<bench>/<task>/

• base__after.mp4: the final program on the unmodified task;
• <vK>__before.mp4: the frozen program on variation vK (expected to fail);
• <vK>__after.mp4: the repaired program on vK (expected to succeed; expected to fail when vK is marked "unsolved": true in index.json);
• index.json: per clip variation, kind, seed_offset, success, pipeline (program version).

Bench	Tasks	Clips
LIBERO (LIBERO-Goal)	open_the_middle_drawer_of_the_cabinet, turn_on_the_stove, push_the_plate_to_the_front_of_the_stove, put_the_wine_bottle_on_the_rack, put_the_cream_cheese_in_the_bowl	53
RoboTwin 2.0	hanging_mug, put_object_cabinet, scan_object, stack_blocks_three, turn_switch	78
ManiSkill3	LiftPegUpright-v1, PegInsertionSide-v1, PlaceSphere-v1, PlugCharger-v1, PullCubeTool-v1	61
RoboDyna	catch_shelf_marble, catch_valley_ball, drop_ball_hole, pick_ripe_apple, place_block_belt	47
RoboPRO	pick_boxdrink_from_basket, put_bowl_in_sink_ks, put_cup_on_coaster, put_mouse_on_pad, put_spoon_on_plate_ks	57

Frame rates: LIBERO clips are real time (5 fps, one frame per 0.2 s of sim time); RoboDyna clips are 5 fps, one frame per 50 physics steps; ManiSkill3 clips are 30 fps, one frame per 2 control steps; RoboTwin clips come from the benchmark's own video recorder or a frame-capture patch for older task backends.

Citation

@inproceedings{li2026a2h,
  title     = {Learning to Anticipate Physical Failures via Agentic Adversarial Harness},
  author    = {Li, Zhiyuan and Liu, Dingsheng and Rasouli, Amir},
  booktitle = {Conference on Robot Learning (CoRL), Late-Breaking Results},
  year      = {2026}
}