FTG64/act-chunking-study

ACT action chunking under delay and disturbance: per-episode results

This repository holds the raw outputs of the study ACT action chunking under delay and disturbance. Every simulated episode behind the study's tables is stored as one JSON record. The repository also holds the scripts that the Hugging Face Jobs ran.

• code/scripts/ holds the scripts as uploaded for the jobs. The GitHub repository is the maintained copy.
• results/<run>/meta.json holds the settings of a run: grid, seeds, deterministic kernels, displacement window, checkpoint, versions and GPU.
• results/<run>/<condition>/seeds_<first>-<last>.json holds one block of 25 seeds for one condition, under the key episodes.

The task is ALOHA transfer cube (gym-aloha), 400 steps at 50 Hz. The checkpoints are the official lerobot/act_aloha_sim_transfer_cube_human ("official") and a retrained FTG64/act_aloha_transfer_cube ("ours").

Runs

Run	Checkpoint	Contents
validation_official	official	Harness vs lerobot-eval on seeds 1000–1049 (harness/, lerobot/), a displacement check, and videos
ckpt_ours_025000 to ckpt_ours_100000	ours at 25k to 100k steps	Full chunk, no delay, seeds 1000–1199
ckpt_official	official	Full chunk, no delay, seeds 1000–1499 (the seeds of its published evaluation)
noise_floor	official	Repeated runs that measure GPU nondeterminism
core_official, core_ours100k	official, ours at 100k	5 execution modes × (5 latencies + 3 displacements), seeds 1000–1199
diag_official, diag_ours100k	official, ours at 100k	Arm trajectories for the first 200 steps (field trajectory), seeds 1000–1049
diag_te_official, diag_te_ours100k	official, ours at 100k	The same trajectories for temporal ensembling: c = 0.01 without latency, c = −0.2 at 0 and 400 ms
te_weights_official, te_weights_ours100k	official, ours at 100k	Temporal ensembling with coefficients 0, −0.05, −0.1 and −0.2, seeds 1000–1199

Condition names

Condition names have the form <mode>_d<delay>_m<displacement>, for example replan25_d00_m6.

• The mode is full (the whole chunk), replan50, replan25 or replan10 (a new chunk every k steps), te (temporal ensembling with ACT's default coefficient 0.01), or te<c> (coefficient c, for example te-0.05).
• The delay is the observation latency in 20 ms control steps: 0, 2, 5, 10 or 20.
• The displacement is the distance in cm that the cube is moved, at a random step inside the run's displacement window (in meta.json; 50–75 for the core, diagnostic and ensembling grids).

Episode fields

Field	Meaning
seed	Environment seed. The displacement direction uses seed + 10000.
success	The cube was transferred to the left gripper (reward 4).
max_reward	Highest reward stage reached: 1 right gripper touches the cube, 2 lifted, 3 left gripper touches, 4 success.
sum_reward	Sum of the per-step rewards.
first_step_at_reward	First step at which each reward stage was reached.
cube_start_xy	Initial cube position in m.
t_move, delta_m, moved_m	Step of the displacement, its vector, and its length in m.
reward_before_move	Reward stage when the cube was moved (0 means not yet touched).
arms_unchanged_by_move	The move left the arm joints unchanged.
mode, k, delay, displacement_cm	The condition's settings. Later runs also record te_coeff.

Load

import glob, json

episodes = [e for f in glob.glob("results/core_ours100k/*/seeds_*.json") for e in json.load(open(f))["episodes"]]