TommyShen/Factor-FB-Offline

FactoredFB Testbed — Expert Demonstrations

Offline expert-demonstration datasets for the Factored-FB testbed, spanning three environment families (Metaworld, ManiSkill, Box2D). Each task ships ~1000 episodes collected with an expert TD-MPC2 policy, stored in a single unified format so one offline loader consumes all of them.

Intended for offline goal-conditioned / factored RL (Forward–Backward representations, FB+FlowBC, GCIQL, CRL, …) and imitation learning.

Families & tasks

Family	Tasks	Steps/ep	Freq	Action	Policy	Filter
Metaworld	mw-assembly, mw-basketball, mw-bin-picking, mw-box-close, mw-button-press	100	40 Hz	4	newt TD-MPC2 (pretrained)	as-collected (~0.9–1.0 success)
ManiSkill	ms-pick-cube, ms-push-cube, ms-stack-cube, ms-poke-cube, ms-lift-peg	25	10 Hz	—	newt TD-MPC2 (pretrained)	as-collected (~0.9–1.0 success)
Box2D (loco)	center-ctrl, goal-ctrl, goal-ctrl-maze, center-wall-navigation	200	—	2	TD-MPC2 (trained from scratch)	success-only
Box2D (manip)	goal-ctrl-target, center-ctrl-target-hard, center-ctrl-target, center-wall	200	—	2	TD-MPC2 (trained from scratch)	success-only

Exact per-task counts, dims, and success rates live in manifest.json.

• Metaworld / ManiSkill policies are the pretrained multitask experts from nicklashansen/newt.
• Box2D policies were trained from scratch (single-task TD-MPC2) on the rlmini_envs Box2D tasks; only successful episodes are kept.

Format

Repository layout:

data/<family>/<task>.npz     # one file per task, arrays stacked over episodes
data/<family>/<task>.meta.json
manifest.json                # global + per-task index/stats

Each <task>.npz holds the 8 unified fields (fixed horizon T per task):

field	shape	dtype
observation, next_observation	[N, T, obs_dim]	float32
action	[N, T, act_dim]	float32
achieved_goal	[N, T, 3]	float32
reward, terminal, timeout, success	[N, T]	float32 / bool

Load

import numpy as np
from huggingface_hub import hf_hub_download

fp = hf_hub_download("TommyShen/Factor-FB-Offline",
                     "data/maniskill/ms-pick-cube.npz", repo_type="dataset")
z = np.load(fp)
obs, act, rew = z["observation"], z["action"], z["reward"]   # [N,T,·], [N,T,·], [N,T]

Policies

The TD-MPC2 policies that collected the Box2D split are published at TommyShen/Factor-FB-Offline-tdmpc2. Metaworld / ManiSkill demos were collected with the public multitask checkpoint nicklashansen/newt.

Helper (list_tasks, load_task) in hf_dataset/load_hf_dataset.py of the Factored-FB repo (branch tl-alignment).

Caveats

• achieved_goal is a placeholder (zeros). FB training does not use it; a goal relabeler fills it downstream. Do not treat it as a real achieved-goal signal.
• The goal is baked into observation for Metaworld/ManiSkill (newt obs), so these are not the classic zero-shot-goal FB observations — factor them out if you need that.
• Box2D reward is potential-shaped (dense, training signal only); the success field is computed from the raw sparse condition and is the authoritative metric.

Provenance / reproduce

Collected with data_collection/collect_expert.py + run_collection*.sh and packed with hf_dataset/pack_and_manifest.py in the Factored-FB repo. Re-run those + push_to_hub.py to extend or refresh the dataset.