Learning to Compose: Recursive Self-Improvement of Shared Executable Stacks for Zero-Shot Long-Horizon Robot Manipulation

Yangzhi Chen; Feiran Liu; Zishuo Wang; Zijian Jin; Xin Zhu; Lecheng Ruan; Wenjuan Han

EMBODIED RECURSIVE SELF-IMPROVEMENT

Recursia.

Learning to Compose

Recursive Self-Improvement of Shared Executable Stacks
for Zero-Shot Long-Horizon Robot Manipulation

Yangzhi Chen¹
Feiran Liu¹
Zishuo Wang¹
Zijian Jin²
Xin Zhu³
Lecheng Ruan⁴
Wenjuan Han^1,5

¹Beijing Jiaotong University
²New York University
³University of Southern California
⁴PKU-HEC joint lab for industrial intelligence, Peking University
⁵LingXing Tech

CoLa Lab

Updated manuscript

Improve one shared executable stack. Validate each repair.
Reuse the frozen stack on unseen task compositions.

Explore task demos Paper PDF Download PDF GitHub

ONE SHARED EXECUTABLE STACKSource-side improvement → frozen target execution

Tools

Reusable perception,
planning & execution

Skills

Programs with observable
execution contracts

Handoffs

Successor readiness
& preserved goals

Verifiers

Object identity
& task relations

Learn from source failuresValidate shared repairsTransfer without target repair

01 / THE IDEA

Skills work alone.
Can they work
together?

Long-horizon robot manipulation requires individually capable skills to compose into physically feasible sequences.

Abstract

Tools can lack a needed capability, report completion too early, or leave a state that the next skill cannot use.

Recursia is a stack-aware, validation-gated framework for recursive self-improvement. Source-task failures guide a coding agent to generate missing tools or repair shared tools, skill contracts, handoffs, and verifiers. Candidates undergo generality checks, mechanism tests, and paired regression tests before acceptance.

The resulting stack is frozen and reused through declarative adapters, making a source-side repair available to multiple skills and unseen compositions. Experiments cover LIBERO-Pro Long simulation and a physical Franka FR3 with Franka Hand. The manuscript reports gains over ASPIRE of 5.8 percentage points on Pos and 15.1 points on Task, and a 40-point mean gain over Skill Document Only across three physical target tasks.

Adapted from the updated manuscript.

02 / THE METHOD

A repair that travels
with every composition.

Experience is retained as shared executable functionality. Each accepted change becomes part of a versioned stack.

Recursia source-side improvement loop: execute and attribute a failure, propose a shared change, validate it, retain or reject it, then compose the frozen stack on target tasks. — Figure 1 Source-task failures drive repairs across tools, skills, handoffs, and verifiers. Candidates pass generality, mechanism, and paired regression checks before entering the shared stack; target execution uses the frozen stack.

Execute & attribute

Run a source task and identify the first supported execution failure.

Propose a shared change

Generate a missing tool or repair a tool, skill contract, handoff, or verifier.

Validate & retain

Check generality and mechanism, then compare paired source outcomes and regressions. Retain or reject.

Freeze & compose

Reuse the accepted stack on unseen target compositions through declarative adapters.

The proposer and evaluator have separate roles. Selection and hidden regression evidence are withheld from the proposer.

Inside the four shared layersInterface examples · Figure 2

Explicit interfaces for composition

Tools: Reusable perception, geometry, planning, and execution operations with explicit inputs and completion signals.
Skills: Parameterized programs with observable prerequisites, completion evidence, and resource limits.
Handoffs: Check what the next skill needs and whether earlier goals remain intact; re-observe, recover within budget, or stop.
Verifiers: Use permitted observations and execution history to assess object identity and task relations.

Contracts provide observable checks, not formal guarantees. In the documented implementation, placement verification runs as shadow monitoring; it does not gate execution or trigger corrections.

Four illustrated interfaces: observed motion termination, placement contracts, grasp-to-transport readiness, and cross-frame object identity. — Figure 2 Four interfaces in one shared executable stack: motion termination, placement contracts, grasp-to-transport handoffs, and cross-frame identity verification.

Completion is not readiness. A predecessor's achieved goal does not guarantee its successor's prerequisites.

03 / SIMULATION TRANSFER

Improve on source tasks.
Compose on unseen targets.

LIBERO-90 → LIBERO-Pro Long.
The shared stack is frozen before target execution.

MANUSCRIPT-REPORTED RESULTS

Recursia target success

28.4%Pos · object positions

53.4%Task · task specifications

Bar charts of LIBERO-Pro Long success under position and task-specification perturbations. Recursia reaches 28.4 percent Pos and 53.4 percent Task in the updated manuscript. — Figure 3 Frozen-stack transfer from LIBERO-90 to LIBERO-Pro Long, reported separately for changes in object positions (Pos) and task specifications (Task). Baselines are reproduced comparisons; execution policies and grasping backends differ.

The manuscript evaluates 10 target tasks with 50 seeds each: 500 episodes per perturbation axis. Rates are task-macro averages. The supplied PDF does not include the referenced per-task tables or uncertainty estimates.

How transfer changes with source experienceSeparate checkpoint study · Figure 7

Success rates at source-improvement checkpoints from N equals 0 through N equals 18; the final shown checkpoint reaches 23.2 percent Pos and 43.4 percent Task. — Figure 7 Separately frozen stacks at successive source-improvement checkpoints. Transfer changes nonmonotonically as source experience accumulates; each checkpoint is evaluated without target-driven stack modification.

N counts improvement iterations, with five source samples per iteration. The N=18 checkpoint reaches 23.2% Pos and 43.4% Task. This is a separate study; its endpoint is not the 28.4% / 53.4% primary configuration above.

04 / REAL-ROBOT DEMONSTRATIONS

Familiar skills.
New compositions.

Three physical target tasks on a Franka FR3 with Franka Hand. Source knowledge stays frozen throughout evaluation.

TASK 01Shared destination

Shared basket

Place the mug and milk carton in the same basket.

1:01.3 · Original speedOpen video ↗

Mug → basketMilk → basket

TASK 02Reassigned goals

Reassigned plates

Place the blue mug on the left plate and the white mug on the right plate, reversing the source assignments.

0:54.5 · Original speedOpen video ↗

Blue → leftWhite → right

TASK 03Dependent subgoals

Bowl–mug dependency

First place the bowl on the plate, then place the mug in the relocated bowl.

0:52.1 · Original speedOpen video ↗

Bowl → plateMug → bowl

Selected successful motion recordings at original speed. Each video includes all supplied motion segments in chronological order; unrecorded perception and model-wait intervals are omitted. Displayed durations are recorded motion time, not full trial wall-clock time.

05 / PRELIMINARY PHYSICAL RESULTS

From source experience
to physical composition.

A first comparison with Skill Document Only (SD-only), which retains source experience in a natural-language skill document.

AVERAGE TARGET SUCCESS

66.7%

Recursia

26.7% SD-only

3 tasks · 10 trials per task per method

Manuscript-reported preliminary physical target-task success: ten trials per task and method.
Target task	SD-only	Recursia
T1 · Shared basket	0.0%	20.0%
T2 · Reassigned plates	60.0%	80.0%
T3 · Bowl–mug dependency	20.0%	100.0%
Average	26.7%	66.7%

Manuscript-reported preliminary physical target-task success on Franka FR3 with Franka Hand (Table 1). Each method runs ten trials per task; success is assessed manually. Average is the mean across the three target tasks. Source knowledge stays frozen, while within-trial feedback can change execution decisions. The selected demo videos are examples, not estimates of success rates.

EVALUATION SETTINGS

LIBERO-Pro Long Franka FR3

06 / ANALYSIS

What makes
composition difficult?

Retrospective cases and a matched inference study examine readiness, validation, and execution on a frozen stack.

RETROSPECTIVE CASE EVIDENCE

A completed goal can still lead to failure.

At N=6, the first mug goal held in 10/10 recorded episodes, but the full mug–pudding task succeeded in 1/10. These are recorded cases, not a controlled ablation.

Two recorded mug-pudding episodes at N equals 6: seed 42 finishes the mug goal but fails to bind the pudding; seed 41 finishes both goals from a different initial state. — Figure 4 Predecessor completion is not successor readiness. In ten recorded N=6 mug-pudding episodes, the mug goal held at termination in all ten, while the full conjunction succeeded once. The illustrated success and failure use different initial states.

SOURCE-SIDE VALIDATION CASES

Local gains must survive paired checks.

At N=9, corridor refusals fell from 8 to 2, but development success improved only from 0/10 to 1/10 and selection stayed at 0/10: the required gains were not met.

Recorded source-side validation examples: a book-transfer repair, a corridor repair that fails required outcome gains, and paired basket-placement regressions hidden by aggregate scores. — Figure 6 A local mechanism repair need not improve full-task completion. Paired regression checks also expose lost successes that aggregate scores can hide.

At N=11, selection improved from 7/50 to 11/50. Yet basket placement had two gains and two losses, exceeding the one-loss per-task cap; the candidate was rejected.

SAME FROZEN N=13 STACK

How much should the driver decide online?

Online decisions match one-shot Pos success and add two percentage points on Task, with higher mean episode time.

Comparison of rule-based, one-shot task-program, and online closed-loop drivers for success rate and mean episode duration on Pos and Task perturbations. — Figure 5 Inference drivers on the same frozen N=13 stack, with 50 matched episodes per perturbation axis. Mean episode times include program generation and robot execution, including failures and timeouts.

50 matched episodes per perturbation axis · manuscript-reported results
Inference driver	Pos success	Task success	Pos mean time	Task mean time
Rule-based	12%	24%	5.28 min	5.20 min
One-shot task program	16%	30%	6.88 min	5.95 min
Online closed-loop	16%	32%	7.22 min	7.15 min

“One-shot” describes the lifetime of the task program. All three drivers may use observation feedback and closed-loop low-level control. Episode means include latency, failures, and timeouts.