faridganbarli/wire_harness_expert_sac

Wire Harness Expert SAC

Expert-policy trajectories collected from the five-mover WireHarness MuJoCo environment for visual world-model training.

Dataset summary

• 20,000 episodes
• 3,491,570 stored observation rows
• At most 300 environment transitions per episode (up to 301 stored rows, including the initial observation)
• 224 x 224 RGB observations, stored as JPEG bytes in pixels
• 10-dimensional continuous actions
• 451-dimensional observations
• Five task stages and sampled start-to-goal transitions

The dataset was collected in 16 independent shards and merged into one Lance table. During the merge, episode indices were regenerated as the contiguous range 0..19999; step_idx starts at zero and is contiguous within every episode.

Validation

The complete merged table was checked locally before upload:

• All 20,000 episode indices are present and contiguous.
• Every episode has contiguous step indices.
• No episode exceeds 301 stored rows.
• Every numeric value in every row was scanned; no NaN or infinite values remain.

See validation.json for machine-readable statistics.

Columns

Column	Type / shape	Description
episode_idx	int32	Globally unique episode index
step_idx	int32	Step index within the episode
state	float32[10]	Controlled mover state
goal_state	float32[10]	Goal state
dynamics	float32[411]	MuJoCo dynamics features
goal_stage	float32[1]	Requested goal stage
start_stage	float32[1]	Initial stage
transition	float32[2]	Start/goal transition pair
stage	float32[1]	Current stage
reset_warmup_ok	float32[1]	Reset warm-up status
render_time	float32[1]	Render timing diagnostic
pixels	binary	JPEG-encoded 224 x 224 RGB frame
observation	float32[451]	Full policy observation
reward	float32[1]	Environment reward
terminated	float32[1]	Termination flag
truncated	float32[1]	Truncation flag
action	float32[10]	Continuous action
id	float32[1]	Environment identifier
variation_task_transition	float32[2]	Sampled task transition

Transition distribution

The first value is represented below as the stage index used by the task: -1 denotes a random initial configuration, while 0..4 denote the five stages. The second value is the goal stage.

Start stage	Goal 0	Goal 1	Goal 2	Goal 3	Goal 4
-1	852	789	875	801	827
0	—	805	799	775	791
1	782	—	807	807	806
2	771	809	—	774	807
3	777	755	809	—	756
4	772	786	826	842	—

Loading directly from Hugging Face

import lancedb

db = lancedb.connect(
    "hf://datasets/faridganbarli/wire_harness_expert_sac"
)
table = db.open_table("wireharness_multi_expert_20k_300")

print(table.count_rows())
print(table.schema)

Example rollout (-1 → 0, random initial configuration to goal stage 0):