Learning from Physical Experience: Context-Space Adaptation for Robotic Manipulation
Learning from Physical Experience:Context-Space Adaptation for Robotic Manipulation
This research introduces LEXP, a framework that enables general-purpose reasoners like ChatGPT-6-Astra to perform physical manipulation tasks by adapting to context through experience. Across 12 real-world manipulation tasks, LEXP achieved 23 out of 36 successful trials (63.9%), significantly outperforming memory-only approaches (52.8%) and vanilla Astra (5.6%). The system combines task-level planning with an experience module that extracts lessons from execution evidence and provides contextual feedback for subsequent actions while keeping model weights and action interfaces fixed. Results show that both task-level planning and retained experience are critical components, as removing either reduces success rates substantially.
The study demonstrates that a fixed-weight general-purpose reasoner can be adapted for robotic manipulation via context-space adaptation, achieving 63.9% success across 12 real-world tasks by combining task-level planning with an experience module that interprets execution evidence and reuses lessons.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
total trials
36
Reported trials
View original evidence
erate contextual feedback for subsequent actions. Model weights and the action interface remain fixed throughout. Across 12 real-world manipulation tasks, LEXP achieves $23/36$ successful trials (63.9\%), compared with $19/36$ (52.8\%) with memory alone and $2/36$ (5.6\%) for vanilla Astra. On three tasks, bread transfer, sort trash and toiletries, and retrieve drawer, removing planning reduces success from $7/9$ to
Open source S4total trials (memory alone)
36
Reported trials
View original evidence
erate contextual feedback for subsequent actions. Model weights and the action interface remain fixed throughout. Across 12 real-world manipulation tasks, LEXP achieves $23/36$ successful trials (63.9\%), compared with $19/36$ (52.8\%) with memory alone and $2/36$ (5.6\%) for vanilla Astra. On three tasks, bread transfer, sort trash and toiletries, and retrieve drawer, removing planning reduces success from $7/9$ to
Open source S4total trials (vanilla Astra)
36
Reported trials
View original evidence
erate contextual feedback for subsequent actions. Model weights and the action interface remain fixed throughout. Across 12 real-world manipulation tasks, LEXP achieves $23/36$ successful trials (63.9\%), compared with $19/36$ (52.8\%) with memory alone and $2/36$ (5.6\%) for vanilla Astra. On three tasks, bread transfer, sort trash and toiletries, and retrieve drawer, removing planning reduces success from $7/9$ to
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2610.41320v1 Announce Type: new Abstract: Recent multimodal foundation models exhibit strong perception, reasoning, and in-context learning capabilities, but translating these capabilities into physical action remains challenging. Through real-world experiments across diverse manipulation tasks, we find that ChatGPT-6-Astra, a state-of-the-art general-purpose reasoner, can serve as a physically actionable, cont
Open source S2
erate contextual feedback for subsequent actions. Model weights and the action interface remain fixed throughout. Across 12 real-world manipulation tasks, LEXP achieves $23/36$ successful trials (63.9\%), compared with $19/36$ (52.8\%) with memory alone and $2/36$ (5.6\%) for vanilla Astra. On three tasks, bread transfer, sort trash and toiletries, and retrieve drawer, removing planning reduces success from $7/9$ to
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org