training_vla_arm: SO-101 sim pick-and-place v3

SO-101 pick-and-place episodes recorded in an Isaac Sim digital twin, packaged as a LeRobot v3.0 dataset for fine-tuning SmolVLA. In each episode the arm picks one pen from the desk and puts it in the pen holder (cup), following a language instruction.

Everything here comes from the simulator. There are no real-robot frames. Read "Known gaps" before training a policy for the real arm.

• repo id: alfinwilliam30/training_vla_arm (private)
• LeRobot codebase v3.0 (written with the team's LeRobot 0.6.2 fork RobotArm_llm), robot_type: so101_follower
• 102 episodes, 139,811 frames at 30 fps (77.7 min of sim time), about 1.3 GB on disk (videos 1.2 GB in 7 files, parquet 6.3 MB, meta 0.8 MB, hil_events 0.3 MB)
• source episodes ep_000282 to ep_000405 of hil_data/episodes, converted by tools/to_lerobot_v3.py (v1.1, --units leader)

Tasks (language instructions)

One task string per episode, chosen by the target pen:

task string	sim object name	episodes
put the blue mechanical pencil in the pen holder	BallpointPen	35
put the red pen in the pen holder	SignPen	36
put the yellow highlighter in the pen holder	Highlighter	31

All three pens lie on the desk in every episode. Only the named pen is moved, so the instruction decides which pen to pick.

Scene variation

• Layout: which pen lies in which of the three desk slots, listed in slot order. The last name is the leftmost pen (robot's left, +y). All six orders appear:

  layout (slot order)	blue pencil	red pen	highlighter
  BallpointPen,Highlighter,SignPen (incl. 18 eps of the "default" scene, ep 282-307)	11	13	11
  BallpointPen,SignPen,Highlighter	4	4	4
  Highlighter,BallpointPen,SignPen	4	4	4
  Highlighter,SignPen,BallpointPen	6	5	4
  SignPen,BallpointPen,Highlighter	5	5	4
  SignPen,Highlighter,BallpointPen	5	5	4

  meta/hil_episodes.jsonl stores the layout as "default" for ep 282-307. Those episodes have the BallpointPen,Highlighter,SignPen order.

• Pose jitter: each pen is moved by up to ±1.5 cm in x/y and ±15° in yaw, from a per-episode seed (seed in meta/hil_episodes.jsonl).

• Pen holder: at world (-0.12, 0.14) m in 91 episodes. In 11 episodes it was moved on purpose (setup.scene.holder_xy in meta/hil_episodes.jsonl): (-0.17, 0.17) m in 6 (ep 328, 329, 362, 363, 384, 385) and (-0.15, 0.58) m in 5 (ep 331, 342, 365, 386, 387). It does not move during an episode.

• Perception: 84 episodes (ep 309+) located the pens with the top-camera perception stack (VLM Set-of-Mark plus a geometry vote). The 18 earlier ones used the sim's ground-truth poses.

Features

key	dtype / shape	meaning
observation.state	float32 (6)	measured joints: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos
action	float32 (6)	commanded joint targets, same names and units
observation.images.top	video 480x640x3, 30 fps	fixed top camera looking down at the desk
observation.images.wrist	video 480x640x3, 30 fps	camera on the gripper
hil.intervening	int64 (1)	1 while a human holds control
hil.control_source	int64 (1)	0 none, 1 policy, 2 voice, 3 teleop
hil.intervention_level	int64 (1)	0 none, 1 stop / yes-no, 2 voice, 3 teleop
hil.phase	int64 (1)	phase of the scripted controller, names in meta/hil_phases.json (idle, open, reach, hesitate, descend, close, lift, carry, lower, release, retreat, home)

The standard LeRobot columns are also present (timestamp, frame_index, episode_index, index, task_index). Video is AV1 (yuv420p). Decode it with video_backend="pyav" if torchcodec is not available.

Units: leader-arm units, the same as the team's so101_sim robot

• observation.state and action use leader units: the numbers the physical SO-101 leader arm reports through LeRobot, and the same numbers --robot.type=so101_sim (RobotArm_llm src/lerobot/robots/so101_sim) accepts and returns. Sim radians were converted with the team's measured linear map isaac_sim/so101-ros2-isaacsim6.1/calibration/so101_leader_to_sim.json (copied as tools/so101_leader_to_sim.json; helper tools/so101_units.py, checked equal to so101_sim's JointMapping to 1e-15).
• So this data can be mixed with teleop recordings made through so101_sim or the leader arm, and a policy trained on it can drive so101_sim directly.
• The sim home pose is about [-9.2, -49.2, 61.7, 37.5, -1.4, 29.2] in these units.
• Older datasets so101_sim_hil_v1/v2 used different numbers (degrees about the URDF zero, home [0, -45, 60, 35, 0, 30], up to ~9 degrees off per joint). Do not mix them with this one without converting.
• action is the target joint position the controller sent at that frame. observation.state is the joint position measured at the same time.

Cameras

• Both cameras are 640x480 RGB at 30 fps, recorded on the sim clock with one frame per control step.
• Top camera pinhole intrinsics: fx = fy = 480 px, cx = 320, cy = 240.
• The wrist camera is attached to the gripper link of the SO-101 URDF (so101_new_calib_camera.urdf).

Human-in-the-loop columns and corrected episodes

• 91 episodes are fully autonomous: the scripted controller ran them with no human input.
• 11 episodes are human-corrected: 286, 288, 289, 293, 294, 296, 298, 299, 301, 304, 305.
  • In each one, an offset of 2 cm left or right was injected into the grasp target on purpose.
  • The policy hesitated, a human gave one voice "nudge" correction (level 2, 1.1 to 2.0 s, 17.4 s in total), and then handed control back.
  • During the correction, hil.intervening = 1 and hil.control_source = 2.
  • These episodes contain a wrong approach followed by a recovery. Leave them out for pure imitation with --dataset.episodes, using the no_intervention list in meta/splits.json.
• The 18 early episodes (ep 282-307) hesitate for 4 to 6 s before grasping (hil.phase = 3). The later ones hesitate for 1 s.

Extra metadata files

file	content
meta/hil_episodes.jsonl	one row per episode: source id, task, layout, seed, outcome, interventions, voice commands, success details, physics report
meta/splits.json	train / val (val is 1 in 10 per pen, stratified), no_intervention, human_corrected, by_pen, by_layout, all as episode indices
meta/selection.json	how the episodes were chosen: rules, kept ids, ids rejected for physics, failed episodes, LLM-planner episodes
meta/physics_check.jsonl	contact and physics report for each kept episode, from tools/physics_check.py
meta/conversion.json	provenance, units and camera intrinsics
hil_events/*.jsonl	the raw event log of each source episode (perception results, grasp plan, voice commands)

How the episodes were selected

The candidates were the source episodes from ep_000282 onward, which use the 16-sided pen meshes and the current physics, up to ep_000405. Selection rules:

1. Policy: setup.policy == "scripted", with a single target pen.
2. Outcome: outcome == "success". The sim checker requires the pen to be lifted and to end up in the holder.
3. Physics check (tools/physics_check.py):
   • pen rise before both jaws touch it ≤ 1 mm
   • release hang ≤ 0.2 s
   • no mid-air pause after release
   • both jaw-to-barrel gaps within ±1.5 mm while the pen is held

Results:

• 102 episodes kept.
• 1 rejected for physics: ep_000364, with a 0.2 s mid-air pause.
• 12 scripted failures not included:
  • 7 misaligned (injected offset, not corrected): 283, 284, 287, 291, 292, 295, 303
  • 1 knocked_holder: 306
  • 1 not_found: 308
  • 3 lifted_not_placed: 325, 330, 340
• 9 LLM-planner episodes are kept out of this single-skill dataset: 332-339 and 343. Their instructions include "책상 정리해줘" and "drop all the things in the cup", and 339 and 343 place all three pens. They are candidates for a later long-horizon (multi-pen) dataset.

Known gaps

• Sim only. Rendered images differ from real camera images in lighting, textures, noise and blur. A policy trained only on this data is not expected to work on the real arm as-is. Plan to co-train with real teleoperated demos (about 50 or more), recorded with the same feature names.
• Grasp assist. While the gripper is closed on a pen, the sim attaches the pen to the jaw with a joint (details.grasp_assist = true in every episode). Holding the pen therefore does not depend on friction, so real grasp robustness has to be learned from real data. The physics filter keeps only episodes where the jaws had actually closed on the barrel before the assist held the pen.
• Scripted demonstrator. Every trajectory comes from one IK-based scripted controller, so the motions are very consistent and show no human style variation. There are also few recovery behaviours: only the 11 corrected episodes contain any.
• Small object and scene variation: three pens, only three holder positions (91 of 102 episodes at the default one), one desk and fixed lighting.
• Units vs a real follower arm: the data is in leader units through the team's so101_sim map. A real follower arm's own calibration was not checked against that map, so verify one known pose before running a sim-trained policy on hardware.
• Some blue-pencil episodes have one jaw touching the desk while closing (jaw_desk_mm ≈ 0 in meta/physics_check.jsonl). ep_000342 has a 6.4 mm sideways push before the grip closes. Both pass the selection rules.

Fine-tuning SmolVLA on this dataset

lerobot/smolvla_base expects three cameras named observation.images.camera1/2/3. Map the two cameras with --rename_map; camera3 is then simply absent. Disable policy.push_to_hub, or give --policy.repo_id.

cd /data/rtxprofuturedrive/alfin/RobotArm_llm
/data/rtxprofuturedrive/alfin/venvs/lerobot/bin/lerobot-train \
  --policy.path=lerobot/smolvla_base \
  --dataset.repo_id=alfinwilliam30/training_vla_arm \
  --dataset.video_backend=pyav \
  --rename_map='{"observation.images.top": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}' \
  --batch_size=64 \
  --steps=20000 \
  --policy.scheduler_decay_steps=20000 \
  --output_dir=outputs/train/smolvla_so101_sim_v3 \
  --job_name=smolvla_so101_sim_v3 \
  --policy.device=cuda \
  --policy.push_to_hub=false \
  --wandb.enable=false

• Local copy without the Hub: add --dataset.root=/path/to/so101_sim_pick_place_v3. Prefix the command with HF_HUB_OFFLINE=1 once lerobot/smolvla_base is in the HF cache.
• LR schedule: scheduler_decay_steps defaults to 30000 in smolvla_base; keep it equal to --steps (RobotArm_llm AGENT_GUIDE.md §7.5). Unfreezing the vision encoder (--policy.freeze_vision_encoder=false --policy.train_expert_only=false, §7.6) is the guide's main recommended change.
• Training subset: add --dataset.episodes='[...]' with the train (or no_intervention) list from meta/splits.json. Alternatively use --dataset.exclude_episodes='[...]'.
• Held-out loss: LeRobot's built-in option is --dataset.eval_split=0.1 --eval_steps=1000. It holds out the last 10 % of episodes per task.
• To push the trained model: use --policy.push_to_hub=true --policy.repo_id=alfinwilliam30/smolvla_so101_sim_v3.
• Inference: use the same task strings, and feed the same two cameras under the same renamed keys.

This command has been tested offline on the local copy: a 3-step run (--steps=3 --batch_size=4, 5 episodes) loaded the dataset, built SmolVLA from lerobot/smolvla_base with the rename map, and trained. The loss was about 0.5 and the run used 2.1 GB of GPU memory.

The upstream SmolVLA docs give about 4 h for 20k steps on one A100 at batch size 64. SmolVLA uses chunk_size = 50 (1.67 s at 30 fps) and the dataset statistics in meta/stats.json for normalisation.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("alfinwilliam30/training_vla_arm", video_backend="pyav")
item = ds[0]   # observation.state (6,), action (6,), observation.images.top (3, 480, 640) float in [0, 1], task