Compile Once, Run Cheap: Frozen, Portable Skills for Embodied Agents
Overview
This research proposes Skill-RSI, a framework for creating reusable manipulation skills for embodied agents through recursive self-improvement. Skills are compiled in simulation and run as standalone code or under a cheap caller, achieving high success rates on new layouts. The study evaluates performance across 7 RoboTwin 2.0 tasks, demonstrating the effectiveness of inheritance over from-scratch rewriting and the cost-efficiency of frozen skills.
From arXiv Robotics — research abstracts
Event evidence and corrections
0 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Environment: simulation
“arXiv:2610.25114v1 Announce Type: new Abstract: Embodied AI lacks the data-rich pretraining that gave language models their agents and skills, so new manipulation capability costs demonstrations per task or a frontier model in every episode. We study a third route: a strong general model acts only as an offline author that writes, runs, diagnoses and repairs a manipulation skill in simulation until it passes an exam”
Exact source · revision 1Source owner not reported
Reported quantity · tasks: 7 other · Basis not reported
“odel family. On 7 RoboTwin 2.0 tasks with cluttered tables, Claude Code compiled 4 tasks into 7 frozen skills, 5 from a first draft needing only qualification, at USD 4.29-8.22 of author cost each (USD 55 per skill counting the 13 failed campaigns). The other three tasks did not compile in 18 campaigns, including five with GPT-6 and a pre-registered 150-layout run: on the hardest task every campaign scored 78-95 of 1”
Exact source · revision 1Source owner not reported
Reported quantity · frozen skills: 4 other · Basis not reported
“odel family. On 7 RoboTwin 2.0 tasks with cluttered tables, Claude Code compiled 4 tasks into 7 frozen skills, 5 from a first draft needing only qualification, at USD 4.29-8.22 of author cost each (USD 55 per skill counting the 13 failed campaigns). The other three tasks did not compile in 18 campaigns, including five with GPT-6 and a pre-registered 150-layout run: on the hardest task every campaign scored 78-95 of 1”
Exact source · revision 1Source owner not reported
Reported quantity · cost per skill: 4.29 other · Basis not reported
“odel family. On 7 RoboTwin 2.0 tasks with cluttered tables, Claude Code compiled 4 tasks into 7 frozen skills, 5 from a first draft needing only qualification, at USD 4.29-8.22 of author cost each (USD 55 per skill counting the 13 failed campaigns). The other three tasks did not compile in 18 campaigns, including five with GPT-6 and a pre-registered 150-layout run: on the hardest task every campaign scored 78-95 of 1”
Exact source · revision 1Source owner not reported
Reported quantity · failed campaigns: 13 other · Basis not reported
“odel family. On 7 RoboTwin 2.0 tasks with cluttered tables, Claude Code compiled 4 tasks into 7 frozen skills, 5 from a first draft needing only qualification, at USD 4.29-8.22 of author cost each (USD 55 per skill counting the 13 failed campaigns). The other three tasks did not compile in 18 campaigns, including five with GPT-6 and a pre-registered 150-layout run: on the hardest task every campaign scored 78-95 of 1”
Exact source · revision 1Source owner not reported
Reported quantity · successes on new layouts: 148 other · Basis not reported
“ampaigns found no detectable first-draft effect of practice (minimum detectable effect 14 points). The frozen skills kept 148/150 successes on new layouts with no model and 147/150 with a DeepSeek-flash caller at USD 0.00028-0.00054 per episode; no direct-control baseline was run.”
Exact source · revision 1Source owner not reported
Reported quantity · cost per episode: 0.00028 other · Basis not reported
“ampaigns found no detectable first-draft effect of practice (minimum detectable effect 14 points). The frozen skills kept 148/150 successes on new layouts with no model and 147/150 with a DeepSeek-flash caller at USD 0.00028-0.00054 per episode; no direct-control baseline was run.”
Exact source · revision 1Source owner not reported
Evidence dependency changes (2)
- 2026-10-30T22:19:14.099Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T22:19:03.402Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
Report timeline
Follow attributed reports and material updates.
- arXiv Robotics — research abstractsSignalCompile Once, Run Cheap: Frozen, Portable Skills for Embodied Agents, and What Recursive Self-Improvement Adds
This research proposes Skill-RSI, a framework for creating reusable manipulation skills for embodied agents through recursive self-improvement. Skills are compiled in simulation and run as standalone code or under a cheap caller, achieving high success rates on new layouts. The study evaluates performance across 7 RoboTwin 2.0 tasks, demonstrating the effectiveness of inheritance over from-scratch rewriting and the cost-efficiency of frozen skills.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).