flower_to_vase_v1: bimanual YAM teleop demos, "Pick up left flower and place into vase"

21 successful teleoperated demonstrations on a bimanual YAM robot (two 6-DoF arms with parallel grippers), recorded 2026-11-11 with raiden's rd record (leader arms). Only the left arm does the task: it picks up the left flower and places it into the vase. The right arm stays at rest.

prompt (every frame's task)	Pick up left flower and place into vase
episodes	21 (the 21 recordings marked success out of 33 attempts; failures are not included)
frames	13,435 at 30 Hz (about 7.5 min, 19-26 s per demo)
start placements	varied freely by the operator from demo to demo (no fixed layout set)
cameras	observation.images.top = scene camera, .left = left wrist, .right = right wrist; 640x360 RGB
observation.state	14 floats, measured: left joints 0-5, left gripper, right joints 0-5, right gripper
action	14 floats, same order: the ABSOLUTE joint targets commanded at that 30 Hz step (radians)
gripper	0 = closed, 1 = open (not normalized)

Folders

• lerobot_v30/flower_to_vase_v1/ is LeRobot v3.0, for current LeRobot (0.4+) and MolmoAct2's LeRobot port. All three cameras.
• lerobot_v21/flower_to_vase_v1/ is LeRobot v2.1, for openpi (pi0 / pi0.5), which pins the older LeRobot. All three cameras.
• train_molmoact2.sh fine-tunes MolmoAct2-BimanualYAM on the v3.0 copy (see below).

Each folder's meta_extra.json records how the frames were built.

How the 30 Hz frames were built

• Source. The robot logs the 30 Hz commanded stream. The cameras are sampled once every 5 control steps (6 Hz).
• Actions. Frame 5t + k (k = 0..4) gets the exact command of control step k after camera frame t. So any 30-frame action window is the real 1-second teleop command stream.
• State. The state is interpolated linearly between the logged measurements.
• Images. Each image is held for the 5 frames, so it is at most 133 ms old.
• Camera orientation. The right wrist camera is mounted upside down. Its frames are turned 180 degrees, the same way the robot serves them at run time.

Download and train

hf download <namespace>/flower_to_vase_v1 --repo-type dataset --local-dir flower_to_vase_v1
cd flower_to_vase_v1

MolmoAct2-BimanualYAM:

1. Install allenai/molmoact2 at commit 6070080, with its lerobot submodule (allenai/lerobot at 80633827, LeRobot 0.5.2).
2. Run the training script:

LEROBOT_DIR=/path/to/molmoact2/lerobot bash train_molmoact2.sh                       # 1 GPU, scene camera only
LEROBOT_DIR=... CAMS=top,left,right N=4 GPUS=0,1,2,3 bash train_molmoact2.sh         # all cameras, 4 GPUs

About the recipe:

• It is our lab's slide-lid recipe: LoRA on the VLM plus the full action expert, 30-step chunks, 5000 steps, batch 32, bf16, and crop / rotation / colour augmentation.
• It needs an 80 GB A100 / H100 class GPU.
• The base checkpoint downloads from the hub on first use (set BASE to use a local copy).

pi0.5 (openpi): train on lerobot_v21/flower_to_vase_v1, using a YAM data config with this layout:

• the 14-D state and action above;
• cameras top / left / right;
• 30-step horizon.

Notes for deployment

• Speed. The demos move faster than the robot's safety layer allows: about half of the 30 Hz steps exceed 0.3 rad/s per joint, and the p95 command speed is about 0.9 rad/s. A policy trained on them imitates that pace. A 0.3 rad/s clamp at run time slows it down, so an episode takes about 1.5x the demo time.
• Start pose. Every episode starts at the robot's home pose with both grippers open.