SimpleICL · In-context Robot Learning Made Simple

02Real-World Experiments

Eight unseen tasks across three difficulty levels.

Evaluation is conducted at three difficulty levels: Easy, where a single executable task is present; Medium, where multiple executable tasks require discrimination; and Hard, where multiple valid execution modes exist.

Level	Method	Weighing	Wipe	Tea	Drawer	Soap	Fruit	Chopsticks	Shelf	Average
Easy
	Fast-WAM	0.90	0.87	0.83	0.80	0.63	0.90	0.73	0.77	0.80
	π_0.5	0.87	0.97	0.80	0.60	0.73	0.60	0.53	0.67	0.72
	SimpleICL (ours)	0.93	0.76	0.73	0.83	0.73	0.86	0.80	0.83	0.81
Medium
	Fast-WAM	0.83	0.50	0.63	0.43	0.60	0.83	0.63	0.73	0.65
	π_0.5	0.63	0.46	0.37	0.43	0.57	0.47	0.23	0.57	0.47
	SimpleICL (ours)	0.83	0.66	0.60	0.80	0.67	0.63	0.73	0.63	0.70
Hard
	Fast-WAM	0.40	0.60	0.27	0.53	0.33	0.57	0.27	0.37	0.42
	π_0.5	0.53	0.33	0.20	0.17	0.47	0.40	0.10	0.30	0.31
	SimpleICL (ours)	0.80	0.67	0.57	0.67	0.63	0.70	0.76	0.60	0.68

Real-world evaluation on eight unseen tasks. Success rates (%) are reported for each method. Wipe = wiping a plate; Tea = serving tea; Drawer = pulling a drawer; Soap = pressing a soap dispenser; Fruit = placing fruit; Chopsticks = organizing chopsticks and bowls; Shelf = organizing items on a shelf.

In-context Robot Learning Made Simple:
A Democratized Recipe for Manipulation Tasks

01Task Execution Demo

02Real-World Experiments