Scientific Autoresearch: Virtual laboratories for physical R&D

Today, we are releasing Dynamical CLI, an open-source execution interface for scientific agents and facilities. Agents use it to compose executable virtual laboratories from admitted capabilities, run campaigns, and route physical requests through a fail-closed authority gate.

We distribute Dynamical through a Codex plugin and portable Agent Skills for Claude Code and other compatible agent harnesses. Scientists can also install the CLI directly from PyPI.

Self-driving laboratories connect robotics, measurements, and model-guided experiment selection. Most systems still begin with a scientific frame chosen by people, including the objective, candidate space, measurements, and decision policy. The next step is infrastructure that can help agents design campaigns and improve later decisions, a progression described in The Past, Present and Future of Self-Driving Laboratories and When Self-Driving Labs Begin to Reason.

Virtual laboratories let scientific agents explore counterfactual experiments before physical execution. They let agents learn instrument behavior, operating limits, workflows, and failure recovery. They also generate model-derived scientific data and verified execution trajectories for evaluation and post-training. Physical facilities then resolve the questions that the virtual environment cannot answer. We call this loop scientific autoresearch. We believe this is a path toward the future of physical R&D.

See the virtual laboratory run

The video pairs a complete recorded agent campaign with a synchronized Isaac replay of the same campaign composition. The left panel shows one counted Claude Opus 5 trajectory.

The agent first compiled the campaign for OpenUSD. Dynamical CLI then recompiled the same composition for Isaac visualization. The 2 target-specific worlds have different hashes. Their geometry and calibration layers are byte-identical.

The campaign returns a model-generated oxygen evolution reaction observation with stated uncertainty. The agent uses that observation to revise its request for a physical experiment. Dynamical returns HOLD because the campaign has no approved physical route.

The Isaac replay visualizes the recorded campaign against the compiled scene. It is not live physical execution or a calibrated twin of the full workstation. No new physical experiment was executed. The video is a derived communication artifact from one of the 144 study trajectories.

How Dynamical works

Dynamical starts with an engineering question and the proof needed to answer it. It finds the scientific capabilities that can produce the required observations, resolves compatible providers, compiles an execution contract, and records the evidence behind each decision. This moves campaign design inside the agent workflow while keeping execution authority outside the agent.

The reusable unit is a scientific capability. Each capability defines inputs, outputs, units, operating limits, uncertainty, failure states, provenance, and execution authority. A capability can bind to a simulator, a predictive instrument model, a read-only facility connection, or an approved physical instrument. The same scientific contract can support virtual and physical routes while keeping their evidence classes separate.

This separation lets an agent compare candidate campaigns without treating all outputs as equivalent. A process simulator can estimate how a procedure behaves. An instrument model can predict a scientific observation. A facility adapter can return a physical measurement only after the facility authorizes and executes the request.

Instrument onboarding begins with manuals, interfaces, calibration records, operating limits, safety controls, and facility authority. Dynamical turns this evidence into an instrument skill, an executable adapter, and a typed action and observation contract. Conformance tests check the candidate integration. The facility decides which capabilities and physical routes to approve.

The instrument-onboarding skill helps instrument teams prepare this evidence. It keeps missing calibration, licensing, safety, and facility information as explicit review items. The Dynamical skill gives scientific agents the corresponding campaign interface.

Dynamical composes capabilities, applies facility policy, routes execution, and records evidence. OpenUSD carries the portable compiled world, including the scene, instrument assets, and campaign composition. NVIDIA Isaac Sim provides the embodied execution layer used for visualization. Scientific instrument models produce observations. One hash-bound trace connects campaign actions, scene state, model-generated observations, sample lineage, validation, and replay. The trace preserves which provider produced each observation and which evidence class it belongs to.

The agent controls the hypothesis, experiment parameters, operation order, analysis, and stopping decision. Dynamical controls provider admission, evidence types, trace integrity, facility policy, and physical authority.

From recorded campaigns to virtual laboratories

Proprio showed that agents can learn instrument procedures from simulator feedback. From 18 identical starting drafts, truthful feedback produced 14 non-regressive repairs. Blind retries produced none. A separate verifier controlled admission, and hardware qualification remained separate.

Dynamical-SDL-1 compiled 1,035 recorded materials experiments into controlled campaigns. Local and held-out forecasts moved in opposite directions in 145 of 293 comparable cases. The study measured whether agents converted acquired physical evidence into better decisions.

Training Scientific Judgment tested whether verified campaign environments could improve that behavior. Post-training raised hypothesis accuracy from 55.2% to 79.3% across 29 held-out environments. The new study connects these ideas through a virtual laboratory that the agent composes from available capabilities. It tests whether the resulting virtual evidence changes a later physical request.

Testing virtual evidence

We used the public FastCat dataset to study a nickel-foam oxygen evolution reaction catalyst. Each agent selected a chemical-bath composition. The provider returned predicted overpotential at 10 mA/cm². The archived physical score was the FastCat E_at_10mA field, an iR-corrected potential at the same current density. Lower values were better.

The provider converts the potential to overpotential by subtracting the common 1.229 V oxygen evolution equilibrium reference. This common offset does not change candidate rankings or pairwise differences. The direct physical analysis below reports the archived potential scale.

The Dynamical arm could inspect capabilities, compose a multi-instrument virtual laboratory, and query one calibrated oxygen evolution reaction provider. The computational-only arm received the same task information without access to that provider.

Each candidate pool created one independent decision problem. The agent could explore the candidates available in that pool, record its scientific state, and submit one final request for a physical experiment. Archived outcomes stayed sealed until scoring.

We tested GPT-5.6 Sol, Claude Opus 5, Qwen 3.8 Max, and DeepSeek V4 Flash 0731. Each model completed both arms across 6 independent candidate pools and 3 repeats. This design produced 144 counted trajectories.

The provider returned overpotential predictions with uncertainty for one fixed output and domain. It refused requests outside its table. The provider passed 9 predeclared admission gates. Its validation mean absolute error was 22.21 mV, interval coverage was 0.963, and top-3 retrieval score was 0.823 against a 0.45 threshold.

Only this bounded provider output passed held-out physical calibration and ranking gates. The other virtual-laboratory outputs were execution feedback, not calibrated predictions of physical outcomes. The physical outcomes came from archived FastCat records that the agents could not inspect.

Measuring agent behavior in the lab

Agents with access to the virtual laboratory acquired different evidence, recorded different beliefs, ran different campaigns, and submitted different physical requests. They kept the provider uncertainty and identified its outputs as model-generated predictions rather than physical measurements.

We then removed one decision-driving observation from 4 Dynamical trajectories. We held the model, task, seed, budget, and all other inputs constant. The intervention changed the final physical request in 3 of the 4 cells.

The predeclared primary score used a reference Gaussian process to estimate how much uncertainty the final physical request would remove after the evidence already acquired. We averaged this score across models and repeats within each pool, then treated the 6 pools as the independent units. The mean paired effect was +0.080 nats in favor of Dynamical. The 90% pool-cluster interval was [-0.081, +0.246], and the exact sign-flip result was p=0.5625. The interval included zero, so the primary endpoint did not establish an improvement in physical experiment requests.

We also computed the requested physical outcome directly from the archived FastCat records. This secondary analysis kept the primary endpoint unchanged.

The mean paired difference was 11.57 mV in favor of Dynamical. The 90% pool-cluster interval was [-1.17, +29.30] mV, and the exact sign-flip result was p=0.3125.

Both analyses used the same trajectories and archived outcomes, so they are 2 measures of one study. Virtual-laboratory access changed agent behavior. Physical-request gains were positive in direction but statistically unresolved.

Provider ranking shaped the result

The provider passed its admission gates, but its accuracy changed across the sealed study pools. Mean absolute error increased from 22.21 mV during validation to 48.41 mV across 54 eligible sealed-pool predictions.

The provider preserved strong within-pool rankings in 3 pools and weak rankings in 3 pools. When its top prediction matched the archived physical best, the mean physical effect was +25.85 mV. When it did not match, the mean effect was -2.71 mV. The difference between these 2 groups had an exact permutation result of p=0.100.

This is an exploratory association. The study did not randomize provider fidelity, so it does not establish provider fidelity as the causal mechanism. The study had only 6 independent pools, and we examined several provider-quality measures.

An agent can use a virtual instrument correctly while weak candidate ranking produces weak scientific guidance. Admission establishes performance on a validation set, while campaign value depends on ranking quality where the agent explores. A useful provider must preserve the comparisons that control the next experiment, not only low average prediction error.

Build with Dynamical

This study covered one dataset, one chemistry, one provider output, 4 models, and 6 independent candidate pools. Of 54 sealed-pool outcomes, 43 rested on one physical run. Where replicates existed, the median within-composition range was 39.5 mV, above the observed 11.57 mV mean paired effect. The study did not test persistent learning, general transfer, or new physical discovery.

Scientists can install Dynamical CLI and use the Dynamical skill to compose and test virtual campaigns. These campaigns support counterfactual exploration and produce verified trajectories and model-generated observations for evaluation and post-training.

Facilities and instrument teams can use the instrument-onboarding skill to contribute evidence-backed capabilities. They retain authority over physical routes, scheduling, safety, and execution.

Virtual laboratories expand what agents can test and learn before scarce physical work. Connected facilities return evidence that improves the models, verifiers, environments, and experiment policies used in later campaigns. Whether better calibrated providers and more independent campaigns can convert behavior change into reliable physical gains remains open.

Appendix. Product and evidence links

• Dynamical CLI repository
• Dynamical CLI on PyPI
• Dynamical skill
• Instrument-onboarding skill
• Release-tagged FastCat calibration record
• Apache License 2.0

Install the CLI runtime.

python -m pip install dynamical-cli

Add the Codex plugin.

codex plugin marketplace add Dynamical-Systems-Research/dynamical-cli --json
codex plugin add dynamical@dynamical-systems-research --json

Claude Code and other Agent Skills-compatible harnesses can install the portable Dynamical skill.

npx skills add Dynamical-Systems-Research/dynamical-cli \
  --skill dynamical --global --copy --yes