BioAgent: Self-Directed Skill Acquisition and Capability Growth for Long-Horizon Laboratory Protocols

Gaojing Zhang; Changwen Chen; Zichen Luo; Jialiang Li; Yuhan Wang; Yifei Xie; Wenzhao Lian

cs.ROOct 30, 2026

BioAgent: Self-Directed Skill Acquisition and Capability Growth for Long-Horizon Laboratory Protocols

Authors: Gaojing Zhang, Changwen Chen, Zichen Luo, Jialiang Li, Yuhan Wang, Yifei Xie, Wenzhao Lian

Organizations: Shanghai Jiao Tong University · University of Sussex

Abstract

Executing an unfamiliar laboratory protocol requires a robot to identify missing capabilities and determine whether each operation remains valid as the experiment progresses. This letter presents BioAgent, an embodied agent that addresses these coupled challenges of capability sufficiency and experimental validity. It matches human-reviewed requirements to validated skills and task packages, selects tool-based implementations for missing behaviors, and requests human teaching when needed during learning. Learned skills undergo sequential task-package validation to check outcomes and compatibility between successive operations. During execution, precondition, outcome, and experimental-state checks govern progression, autonomous recovery, or termination. Execution evidence guides tool selection, while accepted experience supports offline policy updates and task packaging. On the second task, BioAgent reuses 17 skill types unchanged, revises two existing implementations, and acquires five new types. On two dual-arm laboratory tasks with 20 and 31 ordered skill invocations, BioAgent achieves complete-task success rates of 66% and 58% with autonomous recovery over 50 runs per task. Across 25 atomic skills evaluated over 50 autonomous trials each with recovery disabled, success rates range from 84% to 100%, with 17 skills achieving 50/50.

Figures & tables

Fig. 2: BioAgent architecture. Protocol requirements guide capability matching, targeted teaching, and tool-based skill implementation. Verified execution produces experience for offline policy learning and task packaging in Stage 5; validated capabilities return to planning in Stage 1. The upper row illustrates the extension from laboratory handling to bimanual and instrument-related operations. Execution records are retained in the experiment notebook.

Module or pipeline	Type	Function in BioAgent
VLA policy inference	Execution module	Predict manipulation actions from visual observations, robot state, and task instructions using $\pi_{0.5}$ [ 17 ] .
Reviewed trajectory replay	Execution backend	Execute a reviewed demonstration trajectory under its specified starting conditions.
Visual alignment and trajectory adaptation	Execution pipeline	Align task-relevant objects with demonstration references and adapt the recorded trajectory to the current scene.
Object-pose-guided trajectory replay	Execution pipeline	Use estimated object poses to adapt and execute a reviewed object-relative trajectory.
Instrument command execution	Device interface	Execute supported instrument command sequences and return execution status.
6D object pose estimation	Perception tool	Estimate the target object’s pose in a specified coordinate frame, e.g., using FoundationPose [ 18 ] .

TABLE I: Tools and tool chains grouped by their primary function.

Fig. 3: Task 1: plate pretreatment and incubator loading on the Monte 02 dual-arm platform. Representative frames show pipette and reagent-tube handling, pretreatment-liquid transfer, and incubator loading, arranged from left to right and top to bottom.

Fig. 4: Task 2: plate washing. Representative frames show incubator retrieval, removal of existing liquid, cleaning-liquid transfer, and wash-liquid removal, interleaved with tube handling and capping-machine operations. Frames are arranged from left to right and top to bottom.

Task	Ordered invocations	Distinct skills	Runs per condition	Without recovery	With recovery
				Complete-task count (SR)
Task 1	20	20	50	27/50 (54%)	33/50 (66%)
Task 2	31	24	50	20/50 (40%)	29/50 (58%)

TABLE II: Task scope and complete-task success with and without recovery (50 runs per task and condition).

		Success rate (%): 50 trials per method and skill
No.	Atomic skill / operation binding	ProtoAct [ 6 ] (SmolVLA)	BioProVLA [ 5 ] (AugSmolVLA)	ENPIRE [ 10 ]	BioAgent
Plate handling
1	Pick up the six-well culture plate from the target bench position	42	66	50	100
2	Retrieve the six-well culture plate from the tray inside the incubator	38	68	48	94
3	Place the six-well culture plate at the target position on the bench	70	82	66	100
4	Place the six-well culture plate onto the tray inside the incubator	36	56	46	100

TABLE III: Atomic-skill success rates (%) over 50 real-robot trials per method and skill on the Monte 02 platform, with recovery disabled.

Fig. 5: VR-based teleoperation setup for collecting human demonstrations for robotic skill acquisition.

Fig. 6: Distribution of the 39 failures in BioAgent’s atomic-skill evaluation (Table III ), with 50 trials per skill and recovery disabled. The other 17 skills succeed in all 50 trials.

Development quantity	Task 1	Task 2
Initial validated skill types	0	20
Required distinct types	20	24
Planner gap flags	20	7
Reference-confirmed implementation gaps	20	7
Unchanged reused types	0	17
Revised existing types	0	2

TABLE IV: Skill acquisition and reuse during task development.

Task 1: 20 invocations
Condition	$k{=}5$	10	15	20	Final
No recovery	38	34	27	21	20
R1 only	38	35	29	24	22
R1+R2	39	37	32	27	26
Task 2: 31 invocations
Condition	$k{=}8$	16	24	31	Final

TABLE V: Task progress and complete-task outcomes in the separate recovery comparison, with 40 runs per task and configuration. Columns indexed by

k

count # of runs completing the first

k

skill invocations; Final counts complete-task successes after terminal checks. All entries are counts out of 40.

Robotic Laboratory Automation

See all 14 papers

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
Oct 1, 2026Yen-Jen Wang, Haozhe Jiang, Shuying Deng +7Robot Skill Learning Large Language Model-Based Robot Planning
UC Berkeley · Amazon FAR · MIT +1
SkillFlow:Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents
Apr 19, 2026Ziao Zhang, Kou Shi, Shiting Huang +13Lifelong Learning Agents Long-Horizon Agent Evaluation
University of Science and Technology of China · University of Toronto · University of Sydney
WetRobo: A Reproducible Robot Kit for Coding Agents in Biological Laboratories
Sep 16, 2026Yuna Oikawa, Kei Endo, Takanori Uzawa +5Robot Policy Adaptation Code Generation for Robotics
Department of Computational Biology and Medical Sciences, The University of Tokyo, Japan · RIKEN Pioneering Research Institute, Japan · New York University, USA +3

Oct 1, 2026J/K move · Enter open · S save

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

Yen-Jen Wang, Haozhe Jiang, Shuying Deng, Haoru Xue+6

UC Berkeley · Amazon FAR · MIT+1

Robot Skill Learning Large Language Model-Based Robot Planning Sim-to-Real Transfer Robotic Manipulation Self-Improving Agents Agentic AI for Robotics

Open paper AskPDF arXiv

Building reliable robot capabilities across diverse tasks requires substantial human effort to develop and maintain skills, design rewards, and integrate perception with control. We present Reconstruct, Practice, Go Real (RPG), a framework for autonomous improvement of robot execution systems without updating model weights. RPG identifies manipulation capabilities in an offline dataset and constructs related practice tasks in simulation. During practice, RPG uses execution feedback, privileged simulator state, and available dataset videos to diagnose failures. It develops new reusable symbolic skills, refines existing skills, and revises the system prompt based on these diagnoses. Cross-task evaluation tests individual candidate changes and merged revisions before they are retained for reuse. At test time, a multimodal LLM uses the resulting system prompt and skill library to coordinate perception and robot control. On held-out initializations of 22 manipulation tasks, RPG improves task success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming all evaluated baselines, including ASPIRE (75.5%) and CaP-Agent0 powered by GPT-6 Astra Pro (60.0%). After a common calibration and hardware-adaptation procedure, the frozen system succeeds in all 30 physical trials, with ten trials on each of three tasks. Project Website: https://rpg-robot.github.io/

Abstract

Robotic Laboratory Automation

Explore similar work

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

SkillFlow:Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents

WetRobo: A Reproducible Robot Kit for Coding Agents in Biological Laboratories