LabEvolver — Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents

Toward automated scientific discovery

From state-valid control to learn-by-doing autonomy

Autonomous laboratories should accumulate episodic memory from concrete operations so that planning, cognition, and action improve together through practice. LabEvolver operationalizes this capability through a nested dual-loop process.

91.4%cumulative Success@20 over 500 continual ALFWorld tasks

95.3%successful locked-parameter pouring validations within ±0.2 g

48.2%lower mean pH-regulation completion time versus ReAct

55.0%less completion time for pH–EC control with full experience

Nested dual-loop framework

State-grounded execution. Experience evolution.

Inner Trial Loop

The Observer builds a hierarchical state from visual, sensor, and actuator feedback. The Operator maps goals to parameterized LabSkill actions, while a tri-layer safety gate checks interface validity, procedural consistency, and physical feasibility before dispatch.

Outer Evolution Loop

The Strategist converts completed execution traces into state-paired skill, strategy, and safety memories. Add, update, upvote, downvote, and forgetting operations keep retrieval useful and the memory bank bounded.

The latest LabEvolver framework: state-grounded execution in the inner loop and cross-experiment experience reuse in the outer loop.

Scalable embodied decision making

Experience improves decisions without memory explosion

Across 500 continual ALFWorld tasks, LabEvolver reaches 91.4% cumulative Success@20—15.2 points above ReAct and 3.8 points above MemP. The advantage generalizes across Claude, GPT, DeepSeek, and Qwen backbones.

ALFWorld cumulative success, cross-backbone results, and memory growth

ALFWorld results. Structured maintenance limits the experience bank to 281 memories after 500 tasks, compared with 1,177 procedural memories for MemP.

Training-free

All gains come from test-time execution experience; model weights remain fixed.

Backbone-agnostic

LabEvolver delivers the highest Success@20 for every evaluated backbone.

Bounded memory

Consolidation and forgetting prevent unbounded context growth while retaining useful experience.

Physical validation

Safe, adaptive execution in the real wet lab

The physical platform combines a RealMan RM75-B arm, EG2-4B gripper, Intel RealSense D435 RGB-D camera, electronic balance, and water-quality meter. We evaluate quantitative pouring, single-objective pH regulation, and coupled pH–EC regulation.

Skill experience

Quantitative pouring

Across water, sucrose solution, and olive oil at 10–40 g targets, 41 of 43 validations fall within ±0.2 g. On 20 g water, LabEvolver achieves 0.083 g MAE in 11.179 s.

Condition-specific experience adapts controller parameters instead of replaying one fixed setting.

Safe inner loop

pH regulation

State feedback drives replanning after every measurement. The safety gate intercepts invalid transitions before physical dispatch, while retrieved experience reduces exploration: a repeated pH 5 task reaches pH 4.88 with only two additions.

Representative pH closed-loop execution with a safety gate warning

Representative closed-loop pH preparation, including an occupancy-conflict interception and state-grounded recovery.

Strategy experience

Coupled pH–EC control

Full task-relevant experience reduces additions from 21 to 11, completion time from 7,947 s to 3,574 s, and gate intercepts from 2 to 0. The system reaches all four evaluated pH–EC target regions.

pH and EC multi-objective regulation results

More relevant experience lowers planning cost while measurement-conditioned replanning handles coupled, nonlinear solution dynamics.

Evaluation	Setting	Key result	Outcome
ALFWorld	500 continual tasks	91.4% cumulative Success@20	Best overall
Pouring	3 liquids × 4 target masses	95.3% within ±0.2 g	Transfers
pH regulation	pH 5 across 3 backbones	2.33 mean additions with LabEvolver	Robust
pH–EC regulation	4 coupled target regions	8 additions on average	4/4 reached

Scope. LabEvolver currently relies on human-designed LabSkills and safety rules. These explicit interfaces make execution interpretable and auditable, but may constrain adaptation to unseen operations and novel failure modes.