LEARN MISSING SKILLS. REUSE VALIDATED CAPABILITIES.

BioAgent

Self-Directed Skill Acquisition and Capability Growth for Long-Horizon Biological Experiments

BioAgent matches a biological protocol to its validated skill library, requests targeted human teaching for missing capabilities, and verifies each operation before continuing. Accepted experience supports offline learning and reuse in subsequent experiments.

▶ Watch Demo View Architecture Project Resources
45+ atomic skill definitionsGrowing protocol repertoire50 trials per evaluated skill
01 WHY BIOAGENT

Unfamiliar protocols.
New skills to acquire.

A robot must identify what it can already do, learn what is missing, and preserve the conditions required by the rest of the experiment.

⟶

Dependent operations

Liquid handling, tube capping, plate manipulation, and incubator access form dependent sequences. Each operation changes the conditions for the next.

⊞

Targeted skill acquisition

BioAgent retains supported operations and refines unsupported requirements into specific learning requests. Human teaching focuses on the missing behavior.

◉

Validity-aware recovery

Retry or recovery depends on the current sample, container, liquid state, and elapsed time. Progress advances only when the required outcome is verified.

02SYSTEM ARCHITECTURE

Analyze. Acquire. Execute. Reuse.

A protocol-driven loop connects capability-aware planning, targeted skill acquisition, tool orchestration, and verified execution.

01

Analyze & review

The Analyzer extracts operations, object bindings, outcomes, and experimental constraints. A human reviews the requirements.

02

Match & refine

The Planner reuses compatible validated skills and task packages, then refines only unsupported requirements.

03

Acquire & realize

Tools provide policies, aligned replay, perception, and device operations. Missing behaviors request targeted teaching.

04

Execute & verify

The Monitor checks readiness and outcomes, then governs continuation, bounded retry, recovery, or termination.

05

Learn & reuse

Accepted demonstrations, execution traces, and corrections support offline policy updates and reusable task packages.

Shared resourcesValidated skills & task packagesExtensible tool catalogExperiment notebook

Notebook entries link the requirement, skill version, observations, executed actions, verified result, interventions, and resulting state.

03REAL-WORLD DEMONSTRATION

BioAgent in the laboratory

Watch long-horizon biological workflows on the Monte02 dual-arm platform, from protocol analysis and targeted teaching to verified execution and skill reuse.

5:00 · 1080p · English narration · Accelerated real-robot footage; playback speeds are marked in the video.

Download full video

JUMP TO A CHAPTER

Select a chapter to play the corresponding part of the demonstration.

04 SKILL REGISTRY

Capabilities as explicit contracts

A skill contract specifies what an operation accepts, when it may run, how it executes, and what evidence confirms success. The compact aspiration example below shows the core schema fields.

45+ atomic skill definitions and growing. The registry spans laboratory manipulation, perception, and navigation, including candidate and draft skills. Each capability requires validation within its declared scope before reuse.

atomic_skill.template.yaml · compact example 
# Compact schema excerpt, illustrated with an aspiration skill.
# Supply the full configuration and validation evidence before execution.
version: 1
type: atomic_skill
skill_id: aspirate_from_tube
kind: pi05_policy
status: needs_validation

info:
  summary: Aspirate the requested liquid from an open tube.
slots:
  targets:
    - {name: tube, object_type: centrifuge_tube, required: true}
preconditions:
  all:
    - {predicate: pipette_has_tip}
    - {predicate: centrifuge_tube_open}
safety:
  forbidden_when: [liquid_spill_detected]
  stop_triggers: [collision_risk]
executor:
  type: monte_pi05
  active_side: left
  checkpoint_path: null # Bind a reviewed checkpoint.
expected_end_state:
  must_hold: [pipette_tip_contains_requested_liquid]
  preferred_views: [head_rgb, right_rgb]
monitor:
  success_predicate: pipette_tip_contains_requested_liquid
  uncertain_policy: stop
effects:
  world_patch:
    relations:
      - {subject: requested_liquid, relation: contained_in, object: pipette_tip}
failure_handling:
  max_retries: 0
  on_failure: human
training:
  validation:
    status: unvalidated
    success_trials: 0
    failure_trials: 0
Compact schema excerpt adapted from BioAgent’s atomic-skill template. Full configuration, parameter coverage, and validation evidence must be supplied before execution.
Download compact template
Identity & bindings

skill_id names the behavior; slots declare its objects. Supported parameter values define the scope of reuse.

Readiness & safety

preconditions define required states; safety lists forbidden conditions and triggers that stop execution.

Execution binding

executor binds the operation to a reviewed policy or tool, its active arm, and the required runtime configuration.

Outcome verification

expected_end_state and monitor specify observable success, camera views, and the response to uncertainty.

Effects & failure handling

effects updates the world only after a successful postcheck. failure_handling limits retries and defines the handoff response.

Validation & promotion

New skills remain needs_validation. training.validation records the evidence required before a reviewed status change enables reuse.

05 MISSION LIFECYCLE

From protocol to a reusable repertoire

Protocol review, capability matching, targeted teaching, verified execution, and offline consolidation connect one experiment to the next.

REVIEW 01

Review

Convert the protocol into ordered, human-reviewed requirements.

MATCH 02

Match

Retain validated skills and packages that satisfy the requested operation and bindings.

ACQUIRE 03

Acquire

Refine remaining gaps and request demonstrations or corrections where needed.

VALIDATE 04

Validate

Test candidate skills and task compositions before registration and reuse.

EXECUTE 05

Execute

Check preconditions, verify outcomes, and recover only when continuation remains valid.

CONSOLIDATE 06

Consolidate

Use accepted experience offline to update policies and construct reusable task packages.

06 HUMAN IN THE LOOP

Teach the missing behavior

A missing aspiration skill triggers a targeted teaching request. Demonstrations and corrections are linked to the requesting operation; accepted data supports an offline policy update, followed by execution validation before reuse.

registry_match: none  →  status=paused   action=request_human_demonstration

24-second sequence · Missing skill → human teaching → offline update → independent execution. Footage is accelerated.

01

Identify the gap

Keep supported operations and isolate the missing aspiration behavior.

02

Request targeted teaching

Collect demonstrations or corrections for the operation, objects, initial conditions, and expected outcome.

03

Update offline

Review accepted observation–action data and refine the compatible policy implementation.

04

Validate & register

Validate the candidate within its declared scope before autonomous execution and subsequent reuse.

skill_registered → available for autonomous reuse

Human demonstrations and corrections belong to the learning stage. Formal evaluation begins after learning and permits no human intervention.

07EXPERIMENTAL RESULTS

From individual skills to complete protocols

Success is measured both at the atomic-skill level and over the full sequence of required experimental outcomes.

25

Evaluated atomic skills

Evaluation subset: plate pretreatment and plate washing, with 20 and 31 ordered skill invocations.

84–100%

Atomic-skill success

50 trials per skill, with recovery disabled.

17/25

Skills achieved 50/50

Separate atomic-skill evaluation on the real robot.

Complete-task success · 50 runs per task and condition
Protocol Skill invocations Without recovery With recovery
Plate pretreatment & incubator loading 20 27/50 (54%) 33/50 (66%)
Plate washing 31 20/50 (40%) 29/50 (58%)

Reported in the BioAgent manuscript, Tables II–III. The evaluated workflows share 19 atomic skills. Atomic skills and complete workflows are evaluated separately. All formal evaluations prohibit human intervention; task-level recovery uses autonomous retries and validated recovery skills.

08EXPERIMENTAL WORKFLOWS

Different protocols. A shared skill repertoire.

BioAgent reuses validated skills across different protocols and acquires missing capabilities as new experimental requirements arise. The workflows below illustrate plate pretreatment, incubator loading, and plate washing. Watch the complete recordings on YouTube.

BioAgent plate pretreatment & incubator loading on the dual-arm robot▶ Watch full Task 1 ↗
Task 1

Plate pretreatment & incubator loading

Prepare a six-well plate, attach a tip, transfer pretreatment reagent, recap the tube, replace the plate lid, and load the plate into the incubator.

20 invocations · 20 distinct skills
BioAgent plate washing on the dual-arm robot▶ Watch full Task 2 ↗
Task 2

Plate washing

Retrieve the plate, remove existing liquid, add cleaning liquid, mix, and remove wash liquid, with pipette-tip changes and tube handling between operations.

31 invocations · 24 distinct skills
PROJECT RESOURCES

Explore the BioAgent project

A research system for acquiring, verifying, and reusing robot capabilities across long-horizon biological experiments.

Code RepositoryComing soon
Explore Skill Contracts Watch Full Demo