You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@@ -30,6 +34,9 @@ export default function Home() {
30
34
<h1className="font-bold text-t1 leading-tight">
31
35
Mastering Garment Manipulation from 0 to 100% in 500,000 rmb/20 Hours
32
36
</h1>
37
+
<i>
38
+
TL;DR: In this blogpost, we investigate an effective pathway to achieve robost manipulation for live-streaming broadcast without heavy resources (compute, data, infra, etc.), just in one day.
39
+
</i>
33
40
<divclassName="flex flex-row gap-20">
34
41
<divclassName="flex flex-col">
35
42
<h2>
@@ -74,8 +81,8 @@ export default function Home() {
Is massive scale the only path to robotic dexterity? While foundational models like Pi05 and GO-1 leverage massive compute and data to achieve task-generalization, we explore a "resource-aware" alternative for state-generalization. We present a comprehensive guideline for mastering long-horizon garment manipulation, taking a system from 0% to 100% reliability with a fraction of the standard cost—specifically, within 20 hours of human demonstration, 8*A100 gpus, rather than tens of thousands of hrs data and hundreds of GPUs which are typically required.
Our system enables collaborative, dual-arm robots to perform complex folding tasks over extended horizons (up to 6 minutes), demonstrating exceptional robustness to the compounding errors that plague collaborative handovers. This reliability extends to "in-the-wild" scenarios, handling interruptions and varied lighting with ease.
We argue that not all data is created equal. By strictly aligning the human demonstration distribution ($P_train$), the model's knowledge ($Q_model$), and the test-time distribution ($P_test$), we minimize in-distribution shifts that lead to failure.
138
+
</li>
139
+
<li>
140
+
<b>
141
+
Model Arithmetic:
142
+
</b>
143
+
Rather than relying on a single checkpoint, we identify that models trained on different data subsets learn distinct manifolds. We introduce a "shortcut" to merge these manifolds—condensing several diverse models into a single, adaptive policy without the complexity of MoE architectures.
144
+
</li>
145
+
<li>
146
+
<b>
147
+
Distribution-Aware Orchestration:
148
+
</b>
149
+
A specialized mechanism to handle the "last mile" of precision (the final 80-100% performance gain).
This recipe is universally applicable, functioning across different embodiments (AgileX Piper, ARX X5) and model architectures (pi series, x-vla, go-1), proving that efficient, high-performance manipulation is accessible without industrial-scale resources.
0 commit comments