galbenecho/echo-foundry-vla

Echo Foundry — VLA Dataset

TL;DR

10,133 episodes (60 h) of physics-verified robot manipulation where every tick explains itself — a reasoning sentence with live measured distances on all 4.3M frames, every object's ground-truth pose, metric depth, typed outcomes with the evidence attached, and the planner's rejected alternatives with the margins that killed them. One robot (UR5e), five tabletop tasks, simulation registered to a real robot cell whose calibration ships inside.

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("galbenecho/echo-foundry-vla", episodes=[0, 1, 2])  # subset: seconds
ds[0]["extra.language.decision"]
# "Lowering the white flask onto the green cube's top center, 215 mm to go."

What you actually get

• 60 hours of manipulation where every frame explains itself. Not captions — computed sentences: "Lowering the white flask onto the green cube's top center, 215 mm to go." The 215 is measured from the same row's state; you can recompute it.
• The failures, kept and diagnosed. 835 failed episodes with the physical reason and the millimetres. 1,057 flagged successes where the flag is a typed object carrying its proof (which gate, which contact, how deep, which frame).
• What the planner didn't do. ~4.4 rejected grasp candidates per episode, each with the reason and the margin that killed it. No collected dataset has this channel — a human demonstrator can't log what they considered.
• The world itself. Every object's ground-truth pose every 50 ms, metric depth per camera, and each episode's initial world as a loadable MuJoCo model.
• A real cell behind it. The simulation is registered to a physical robot cell; its calibration, measured depth noise, and actual room ship in calibration/.

Start here → USAGE.md — seven copy-paste recipes and the map of what links to what. Ten minutes to productive.

Why this exists

every teleop dataset	this dataset
(image → action)	(image → considered → rejected-because → chosen → action → measured outcome)
success demos	failures kept: typed reason + millimetres
captions guessed by a VLM	reasoning text computed from world state — re-derivable from the same row

A human demonstrator can't log what they considered and didn't do. A solver can — and this corpus was made by one that searched, gated, and measured everything it touched. Mean 4.4 rejected grasp candidates per episode, each with the physical reason and the margin in mm.

What's in the box

episodes / frames	10,133 / 4,316,292 @ 20 fps · train/val split cut at family boundaries
cameras	3 × 848×480 RGB + aligned metric depth per camera
per tick	reasoning sentence · skill · held objects · joints+gripper state/action · EE pose · every object's ground-truth pose
per episode	outcome (+evidence flags) · full physics verdict · rejected grasp candidates · materials & masses · camera intrinsics · initial-world hash
worlds/	models/<family_id>.xml — every unique initial world as standalone MuJoCo; shared meshes/textures in assets/
calibration/	the real cell's intrinsics, distortion, camera poses, measured per-pixel depth noise, and the actual room backplates behind every frame
task text	verified natural-language instruction on 95.2% of episodes

The five tasks

Each scene contains ONLY what its row lists — there is no scene with four cubes and a flask. Structures below are read from the shipped placement records (all 132 distinct sequences across 10,133 episodes enumerated — nothing else occurs).

task	scene contains	final structure	what it tests
stack_all	4 cubes	one 4-cube tower ▢▢▢▢	sequential construction — each placement becomes the next one's support
towers	4 cubes	two 2-cube towers ▢▢ ▢▢	same objects, different goal arrangement
flask_mid	2 cubes + flask	▢/⌂/▢ — cube, flask on it, cube on the flask	the support surface changes mid-task; the corpus's hardest family
cube_on_flask	1 cube + flask	▢/⌂ — cube seated on the flask	placement onto non-cubic geometry, anchor-derived
flask_on_cube	1 cube + flask	⌂/▢ — flask seated on the single cube	the inverse support relation

Outcomes: three tiers, evidence attached

A success tier requires achievement evidence — recorded, seated placements. next.success=True only on clean_success.

outcome	n
clean_success	8,241 (81.3%)	all placements seated, nothing flagged
flagged_success	1,057 (10.4%)	goal achieved; flags below say what happened
failed	835 (8.2%)	goal missed — with the why and the millimetres

outcome_flags are typed objects carrying their proof (severity 3→1):

[{"kind": "gate_refused", "severity": 3,
  "search_refusals": {"flight_margin": 13, "arm_collision": 8, "workspace": 1},
  "worst_margin_mm": {"flight_margin": 10.8, "workspace": 70.0}},
 {"kind": "unplanned_contact", "severity": 2,
  "contacts": [{"a": "work_table", "b": "ur5e_robotiq", "depth_mm": 0.4,
                "frame": 142, "phase": "in_transit"}]},
 {"kind": "bystander_disturbed", "severity": 1,
  "disturbed": [{"slug": "cube_3", "moved_mm": 29.8}]}]

Gate glossary — what each margin number measures

gate	checks	the number is
workspace	pose outside the arm's reach	shortfall, mm
flight_margin	carried object keeps comfort clearance	measured min clearance, mm
arm_collision / payload_collision / self_collision	swept-volume penetration	depth, mm
keepout	motion enters a keep-out zone	intrusion, mm
grasp_not_earned	pads never force-closed	per grasp gate
held_flip	wrist would flip IK branch while holding	per its gate

Sim-to-real

Built for the zero-shot experiment: train here, evaluate on the physical cell this data is registered to — its calibration, depth noise, and room ship in calibration/. Real-trial measures are the same columns the synthetic verdicts carry, so sim and real compare column-for-column. Real-robot results not yet included; when published they'll use this protocol.

Format details (depth codec · worlds · splits · loading)

Layout — LeRobot v3.0, chunked ≤1000 files/dir. Full download ~117 GB; use episodes=[…] for subsets.

Depth — depth/observation.depth.<cam>/chunk-CCC/file-EEE.u16seq: uint16 millimetres, LZMA. Frame count = episode length + 1 (terminal observation). Foreground = render ground truth; background = the real room's measured depth (dropouts = 0, ~79% valid). Reader (stdlib only):

import lzma, numpy as np
b = open(path, 'rb').read()
nl = b.index(b'\n'); n, h, w = (int(v) for v in b[:nl].split())
depth = np.frombuffer(lzma.decompress(b[nl+1:]), dtype=np.uint16).reshape(n, h, w)

Worlds — worlds/models/<family_id>.xml: the episode's initial world as a standalone MuJoCo model (shared meshes/textures in worlds/assets/). extra.kernel.initial_state_sha1 is the identity key of the authored source world — episodes sharing it started byte-identical; source-level verification available on request.

Splits — info.json declares train/val cut at a counterfactual-family boundary: sibling episodes share a byte-identical initial world; a row-level split would leak starts across sides.

Companion — echo-foundry-world regroups a subset into counterfactual families (same start, different action order, divergent outcomes). The sets overlap — don't train on one and evaluate on the other.

Provenance

Authored worlds → solved plans → physics rollouts → rendered film, composited over the real cell's captured room. Full corpus 100.7 authored hours at a measured $0.49 per motion-hour on one RTX 3090 — scaling is a compute purchase, not a collection program.

Contact: Echo Labs — sapir@echo-robots.com