PRISM is a research framework for automated experimental protocol generation, validation, and execution in robotic laboratories. It integrates language-model–based reasoning, simulation-driven validation, and robot-aware execution to enable end-to-end automation without human intervention between experimental steps.
This repository accompanies the PRISM paper and provides the prompts and code used for protocol planning, protocol generation, simulation-based validation, and robotic execution.
Figure:Overview of the PRISM framework for protocol generation and execution.
The system consists of three main stages: Protocol Planning, where user intent is converted into structured steps; Protocol Generation, where structured English instructions are transformed into robot-aware actions and iteratively refined through validation cycles in Omniverse before execution; and Real-World Execution, where the full pipeline is validated using the Luna qPCR protocol in our autonomous laboratory.
PRISM operates as a closed-loop system with three core stages:
User intent is converted into structured natural-language experimental steps using language-model–based reasoning.
This stage may involve:
- Automatically retrieving reference procedures from web-based sources
- Generating structured experimental steps (e.g., liquid handling, timing, dependencies)
- Identifying required reagents, instruments, and constraints
Structured protocol descriptions are transformed into robot-aware, executable protocols.
This stage includes:
- Translation into the Argonne MADSci protocol format
- Coordination across multiple robotic instruments
- Simulation-based validation in a digital twin environment built in NVIDIA Omniverse
- Iterative refinement cycles where detected physical or sequencing errors are reported back and corrected
Protocols must pass simulation-based validation before execution.
Validated protocols are executed on an autonomous laboratory platform composed of off-the-shelf robotic instruments, including:
- Opentrons OT-2 liquid handler
- PF400 robotic arm
- Azenta plate sealer and peeler
The full pipeline is demonstrated using Luna qPCR amplification.
PRISM supports systematic benchmarking across:
- Single-agent vs multi-agent protocol generation
- Constrained vs open-ended prompting paradigms
- Protocol correctness, ordering, and refinement efficiency
Simulation-based validation enables consistent detection of physical infeasibility prior to real-world execution.
PRISM/
├── run_prism.py # End-to-end pipeline (Stage 1 → Stage 2)
├── ProtocolPlanner/ # Stage 1: Protocol Planning
│ ├── run_stage1.py # Multi-agent / single-agent LLM pipeline
│ ├── requirements.txt # Python dependencies (openai, anthropic, google-generativeai)
│ └── Prompts/ # Prompt templates per experiment and paradigm
│ ├── PCR/ # PCR: constrained, open-ended, single-agent variants
│ └── CellPainting/ # Cell Painting: multi-agent and single-agent
├── ProtocolGenerator/Code/ # Stage 2: Protocol Generation + Simulation Validation
│ ├── run_agent.sh # Launches Claude Code for autonomous protocol generation
│ ├── projects/prism/ # PCR project: prompts, workflow configs, simulation launcher
│ ├── slcore/ # Simulation core (robot servers, REST gateway, motion)
│ └── assets/ # 3D robot models and labware (USD format)
└── outputs/ # End-to-end pipeline outputs (auto-generated)
Stage 1 (Protocol Planning):
- Python 3.10+
- API key(s) in
ProtocolPlanner/.env:OPENAI_API_KEY,ANTHROPIC_API_KEY, and/orGOOGLE_API_KEY
pip install -r ProtocolPlanner/requirements.txtStage 2 (Protocol Generation + Simulation):
- Linux with NVIDIA GPU (Isaac Sim 5.1)
- Docker and Docker Compose (for MADSci services)
- Claude Code CLI (
claude) - See
ProtocolGenerator/Code/README.mdfor full setup
# Full pipeline: Stage 1 (LLM planning) → Stage 2 (code generation + simulation)
python run_prism.py --experiment pcr --paradigm constrained --model gpt-5
# With Claude Opus, open-ended paradigm
python run_prism.py --experiment pcr --paradigm open-ended --model claude-opus# Run protocol planning, skip simulation
python run_prism.py --experiment pcr --paradigm constrained --model claude-opus --stage1-only
# Or call Stage 1 directly
python ProtocolPlanner/run_stage1.py --experiment pcr --paradigm constrained --model gpt-5
# List all available configurations
python ProtocolPlanner/run_stage1.py --list# Point Stage 2 at a previously generated protocol
python run_prism.py --experiment pcr \
--stage2-only ProtocolPlanner/outputs/pcr_constrained_gpt-5_20260323_120000/final_protocol.txt| Name | Provider | Model ID |
|---|---|---|
gpt-5 | OpenAI | gpt-5 |
gpt-4o | OpenAI | gpt-4o |
claude-opus | Anthropic | claude-opus-4-6 |
claude-sonnet | Anthropic | claude-sonnet-4-6 |
gemini-pro | gemini-2.5-pro | |
gemini-flash | gemini-2.5-flash | |
gemini-flash-lite | gemini-2.5-flash-lite |
PRISM supports systematic benchmarking across:
- Single-agent vs multi-agent protocol generation
- Constrained vs open-ended prompting paradigms
- Protocol correctness, ordering, and refinement efficiency
Simulation-based validation enables consistent detection of physical infeasibility prior to real-world execution.
This repository is intended for research and benchmarking purposes. Protocols generated by PRISM should be independently reviewed and validated before use in safety-critical or production laboratory environments.
If you use PRISM or build upon this work, please cite:
@article{hsu2026prism,
title = {PRISM: Protocol Refinement through Intelligent Simulation Modeling},
author = {Hsu, Brian and Setty, Priyanka V. and Butler, Rory M. and Lewis, Ryan and Stone, Casey and Weinberg, Rebecca and Brettin, Thomas and Stevens, Rick and Foster, Ian and Ramanathan, Arvind},
journal = {Digital Discovery},
year = {2026},
publisher = {Royal Society of Chemistry},
doi = {10.1039/XXXXXXXXX}
}