williamTLmiller/robot-learning-for-beginners-robot-attempts-explained

Robot Learning for Beginners: Robot Attempts Explained in Plain English (CSV)

Never worked with robot data before? Start here. Every row of robot_attempts_simple.csv is one try by a real or simulated robot at a task, described in plain English. This public repo is a free toy sample: 1,000 attempts by 35 kinds of robots from 95 open datasets, spread over all ten collections. No special software: open it in Excel, Google Sheets or pandas.

Free toy sample

This public repo is a deliberately tiny, CSV-only toy sample: 1,000 robot attempts plus the rows linked to them. It is meant for looking at the schema and running a pipeline once (a toy training run), not for training a real model. The full and training-ready versions (williamTLmiller/robot-learning-for-beginners-robot-attempts-explained-full) are private and licensed separately; contact energea@energeatech.com.

The 5 words you need

word	meaning
attempt (episode)	one try by a robot at a task, from start to finish
step (frame)	one moment inside an attempt; robots record many per second
state	numbers describing the robot's body at that moment (joint angles, gripper opening...)
action (command)	the numbers sent to the motors at that moment
policy	the program (often a neural network) that looks at the state and instruction and picks the action. Robot datasets exist to teach policies.

Files

file	rows	what it is
robot_attempts_simple.csv	1,000	one row per attempt: robot, task, how long, how much it moved, success (if recorded), train/test
robots.csv	35	every robot in these attempts, what type it is, how many datasets, attempts and steps it has in the full release
datasets.csv	95	every original dataset these attempts come from, its robot, what each state/command number means, license and link (attempt and step counts are for the full release)
walkthroughs/*.csv	1,130 in 10 files	one attempt per collection, one row per moment, every joint in its own named column
data_dictionary.csv	50	every column explained

Walk-throughs (read one robot attempt moment by moment)

Each walk-through is one of the 1,000 attempts in robot_attempts_simple.csv.

file	robot	instruction
walkthroughs/bimanual-two-arm-robot-manipulation-aloha-demonstrations.csv	ALOHA (2x ViperX 300)	"Slide open the ziploc bag."
walkthroughs/humanoid-robot-demonstrations-unitree-h1.csv	Unitree H1	"Greet the other robot with a high five."
walkthroughs/robot-simulation-benchmark-trajectories-libero-aloha-sim-pusht.csv	PushT pusher (simulated)	"Push the T-shaped block onto the T-shaped target."
walkthroughs/robocasa-simulated-kitchen-robot-demonstrations.csv	RoboCasa kitchen robot (simulated)	"SlideDishwasherRack"
walkthroughs/low-cost-robot-arm-demonstrations-so100-so101-koch.csv	SO-101	"pink lego brick into the transparent box"
walkthroughs/industrial-robot-arm-assembly-insertion-ur5-fanuc-kuka.csv	KUKA iiwa	"insert the peg into the hole"
walkthroughs/kitchen-household-robot-manipulation-real-world.csv	DLR SARA	"Pour into the mug"
walkthroughs/tabletop-robot-pick-and-place-language-instructions.csv	Sawyer	"pick up kettle "
walkthroughs/mobile-robot-navigation-and-mobile-manipulation.csv	Hello Robot Stretch	"pull open a dishwasher"
walkthroughs/cloth-folding-cable-routing-deformable-robot-manipulation.csv	Franka	"route cable"

Try it

import pandas as pd
base = "https://huggingface.co/datasets/williamTLmiller/robot-learning-for-beginners-robot-attempts-explained/resolve/main/"
df = pd.read_csv(base + "robot_attempts_simple.csv")

df["robot_type"].value_counts()                                  # what kinds of robots are there?
df.groupby("robot")["seconds_long"].median().sort_values()       # which robots do the longest tasks?
df[df["did_it_succeed"] != "not recorded"]["did_it_succeed"].value_counts()   # success where recorded
df[df["what_the_robot_was_asked_to_do"].str.contains("fold", case=False)]    # every folding attempt here

First machine-learning project (a toy run): predict task_type from the motion columns (seconds_long, how_much_the_robot_moved, how_jerky_the_commands_were, percent_of_time_holding_still, numbers_in_each_motor_command). Train on used_for == "train" rows, test on used_for == "test". Attempts never appear in both, so your score is honest. This toy sample has 971 train, 14 validation and 15 test rows, so the test score is only a rough check.

Next step: open a walk-through file and plot a joint over time, then move on to the step-by-step data in any collection below.

What is in this toy sample

• 1,000 attempts from 95 datasets (10 to 11 attempts per dataset) and 35 kinds of robots
• robot types: one arm (547), two arms (234), mobile base with arms (114), humanoid (40), wheeled mobile robot (33), four-legged robot (21), simulated 2D pusher (11)
• real or simulated: real (850), simulation (150)
• task types: pick and place (331), open or close (138), insert or assemble (95), other (84), fold (61), push or slide (61), cook or stir (36), navigate (36), hand over or transfer (30), press a button (28), wipe or clean (28), pour (22), route a cable or hose (18), not stated (11), greet or gesture (10), stack (7), sort (4)
• 11 attempts have (no instruction recorded)
• did it succeed: not recorded (978), no (12), yes (10)
• used for: train (971), validation (14), test (15)
• picked deterministically: the ten walk-through attempts and the attempts in the ten collection toy samples first, then attempts taken in turn from each dataset in a fixed hash order

The collections (each has its own free toy sample with a few attempts step by step)

collection	attempts here	attempts in the full collection	what's in it
Two-Arm (Bimanual) Robot Manipulation Demonstrations: ALOHA, xArm and YAM	172	2,502	two-arm robots doing real tasks like coffee pods, taping, threading velcro and towels
Humanoid Robot Demonstrations: Unitree H1 Folding, Warehouse Picking, Rearranging and Greeting	40	122	a full-size humanoid robot folding clothes, picking in a warehouse, rearranging objects and greeting
Robot Simulation Benchmark Trajectories: LIBERO, Meta-World MT50, VLABench, RoboMME, ALOHA Sim, PushT	139	20,376	simulated robots from the most-used robot-learning benchmarks
RoboCasa Simulated Kitchen Robot Demonstrations: Human-Teleoperated Kitchen Tasks	11	25,307	a simulated home robot doing kitchen chores (doors, drawers, appliances, pick and place)
Low-Cost Robot Arm Demonstrations: SO-100, SO-101 and Koch Pick-and-Place, Sorting and Stacking	82	510	cheap 3D-printed robot arms anyone can build, picking, sorting and stacking
Industrial Robot Arm Demonstrations: Assembly, Peg Insertion and Clamping on UR5, Fanuc, KUKA, Franka	96	7,253	factory-style arms doing assembly, insertion, clamping and tabletop work
Real-World Kitchen and Household Robot Manipulation: Pouring, Cooking, Fridge Doors and Play Kitchens	129	11,357	real robots in kitchens and homes: pouring, cooking, opening fridges, tidying
Tabletop Robot Pick-and-Place with Language Instructions: Franka, xArm, Sawyer and PR2	151	7,035	robot arms picking, placing and pushing objects on a table, often with a typed instruction
Mobile Robot Navigation and Mobile Manipulation: Jackal, TurtleBot, Stretch, Go1 and Mobile ALOHA	125	23,037	robots that drive, walk or roll around: indoor/outdoor navigation, doors, elevators, cabinets
Cloth Folding, Cable Routing and Hose Handling: Deformable-Object Robot Manipulation	55	10,216	robots folding clothes, routing cables and moving hoses (soft, floppy objects)

The full release (private)

Everything in this section describes the full private release, not the files in this repo.

The full beginner set, williamTLmiller/robot-learning-for-beginners-robot-attempts-explained-full, has one row for every attempt of every collection: 107,715 attempts by 35 kinds of robots from 95 open datasets. The columns ending in _in_full_release in robots.csv and datasets.csv count that full set. The full collections (every recorded step) are private too. Contact energea@energeatech.com.

Honest labels

Instructions and success are copied from the original datasets, never invented: (no instruction recorded) and not recorded mean the source did not have them. task_type is a simple keyword label made by machine. Robot names come from the source metadata or the source project's documentation (not stated if neither says).

License

CC BY-NC 4.0 (non-commercial) for this packaging; see NOTICE for what that does and does not cover. The original datasets keep their own open licenses (apache-2.0, cc-by-4.0, mit) — links are in datasets.csv; please cite them. Commercial licensing, the full set of attempts and the full collections: energea@energeatech.com.