Dr.Anmar

A clinician-centered surgical robotics simulation-training platform for teleoperation, data collection, and policy evaluation.

Dr.Anmar integrates interactive surgical workspaces, articulated robot systems, OpenUSD assets, PhysX mechanics, multimodal sensing, demonstration recording, and reproducible experiment contracts. The platform is designed to connect a clinician's procedural intent to inspectable robot behavior without hiding the underlying simulator, control, data, or validation boundaries.

Dr.Anmar owns the doctor-facing workflow, procedure rooms, interaction contracts, safety controls, evidence pipeline, and study lifecycle. NVIDIA Isaac Sim, Isaac Lab, PhysX, ORBIT-Surgical-derived task foundations, and optional providers execute bounded technical roles. The complete ownership and provenance model is documented in docs/OWNERSHIP.md.

Caution

Dr.Anmar is software for simulation training, synthetic data, and evaluation. Real-world and clinical evidence are not established. It is not a medical device and must not be used for diagnosis, treatment, patient-specific planning, or control of physical surgical hardware.

Product scope

Dr.Anmar supports four connected activities:

1. Simulation: compose articulated instruments, anatomy, sensors, contacts, deformables, particles, and task-specific mechanics in versioned OpenUSD scenes.
2. Human demonstration: control one or two instruments by keyboard, game controller, voice, or camera-native hand tracking while preserving audited command and simulator state.
3. Robot learning: record synchronized observations and actions, construct dataset cards, train bounded imitation- or reinforcement-learning experiments, and compare policies under controlled perturbations.
4. Evaluation: reproduce task phases and failures, measure native simulator outcomes, retain provenance, and distinguish repository verification and native-simulator evidence from real-world, clinical, and regulatory evidence.

The platform is intended for reproducible engineering studies. Every quantitative result must be interpreted within the exact asset revision, software stack, hardware, scenario, and measurement contract that produced it.

Executable Autonomous Rescue policy loop

Autonomous Rescue OR has a closed imitation-learning loop: record contact-driven expert episodes, pack whole episodes, train Robomimic BC, load one immutable checkpoint into the live room, and evaluate seeded rollouts against patient effects. The policy can emit bounded robot actions only. Bilateral contact, compression, perfusion, overload, vital signs, release, and success remain post-physics outputs of the patient runtime.

./dr_anmar_rescue_il.sh policy-room /path/to/model_epoch_200.pth 2361
./dr_anmar_rescue_il.sh rollout 2361
./dr_anmar_rescue_il.sh evaluate-policy 2361 20 --continue-on-error

Every policy rollout is recorded with its checkpoint SHA-256 and patient-effect outcome. It is labeled learned_policy_rollout and is never automatically accepted as an expert Behavior Cloning reference.

Asset catalog and reproducibility

Dr.Anmar uses one provider-aware asset registry for repository assets and the pinned NVIDIA Isaac for Healthcare v0.7.0 catalog. It inventories complete asset directories, validates relative USD dependency closure, rejects path traversal and workstation-absolute references, checks manifests and licensing evidence, and can generate deterministic directory hashes for release locks. The hub, workstation, native-room resolver, installers, and capability API use the same provider-relative paths instead of maintaining independent root logic. The checked-in lock and generated catalog.md cover all local asset units and all 20 product portfolio entries; CI rejects any drift between those release artifacts and the repository.

python3 scripts/dr_anmar_asset_registry.py verify
python3 scripts/dr_anmar_asset_registry.py inventory --hash
python3 scripts/dr_anmar_i4h_receipt.py verify

The clinician-facing capability payload is generated from all 20 entries in physics_next/dr-anmar-assets.json, including every declared profile, runtime, report, native-evidence, and composition artifact. Repository verification is not a substitute for native-simulator, real-world, biomechanical, or clinical evidence. See docs/ASSET_CATALOG.md.

Dr.Anmar does not use “validated” as a catch-all. Product capability, repository verification, native-simulator evidence, real-world evidence, and clinical evidence are separate claims. See docs/EVIDENCE_LEVELS.md.

Dr.Anmar robot systems

Seven procedure-specific systems are available as simulation-training workcells. Native-simulator runs were recorded on 25 July 2026 for the exact revisions and stacks named in their evidence artifacts. Each system provides:

• a standalone articulated mechanism for isolated development;
• a composable payload mounted to an Isaac Lab Franka articulation;
• deterministic source generation and asset manifests;
• OpenUSD, GLB, texture, interaction-frame, controller, and task contracts;
• CPU-side structural and controller tests; and
• a headless CUDA native-simulator evidence program.

The animations below were rendered directly from the complete Franka-mounted OpenUSD assemblies in Isaac Lab on an NVIDIA RTX 4090. Each clip opens on the full robot, then moves to the authored tool center point and procedure fixture while the mechanism executes its phase targets. They are simulation visualizations, not physical-performance or clinical evidence.

Wound preparation robot

An articulated concentric work head combines a compliant contact guard, interchangeable debridement cartridge, multi-nozzle irrigation, annular aspiration, and explicit fluid-volume accounting. The simulation model represents adhered debris release through accumulated contact work and conserves emitted, active, aspirated, spilled, and discarded particle volume.

These three clips come from the same fresh Isaac Lab run. The first preserves the complete Franka workcell context; the two shorter clips retain the simulated frames while focusing the contact/debridement and irrigation/aspiration portions of the sequence.

The recorded CUDA run covered both standalone and Franka-mounted representations for 120 steps, all five tool joints, a cooked surface-deformable wound, seven debris attachments, 80 PBD particles, zero fluid-ledger balance error, finite joint state, and zero error-level engine messages. See docs/VALIDATION.md.

Atraumatic exposure robot

The exposure system uses symmetric carriages, independent lift and pitch axes, compliant pad travel, and twelve distributed capture cells. Fenestrated and microcup pad variants share one articulation and force/visibility control contract, enabling controlled comparison of contact geometry without changing the experimental interface.

Both pad geometries passed the standalone and Franka-mounted 120-step CUDA matrix with finite articulation state, two cooked tissue flaps, two outer anchors, twelve capture constraints, finite controller output, and zero error-level engine messages. See docs/atraumatic_exposure_robot/VALIDATION.md.

Adaptive hemostasis robot

This system combines bilateral compression, irrigation and annular suction, clip delivery, patch application, and a reduced-order pressure/flow verification model. Its runtime contract separates temporary compression, retained-clip, and patch-bond attachments so that control phases and failure conditions remain independently inspectable.

Qualification checks current surface-deformable vessel schemas, all eleven tool joints, attachment lifecycles, conserved particle-volume bookkeeping, suction capture, provisional retention/cure thresholds, pressure-challenge integration, finite state, and error-free engine execution in standalone and Franka-mounted configurations. See docs/adaptive_hemostasis_robot/VALIDATION.md.

Adaptive anastomosis robot

The anastomosis system provides bilateral circumferential capture, coaxial tissue approximation, an expandable lumen mandrel, independent eversion, a sixteen-position staple crown, reinforcement-collar application, temporary occlusion, and a pressure-decay test model.

The native runtime matrix covers the 14-DoF standalone mechanism and 21-DoF Franka assembly, two cooked tissue surfaces, twelve temporary capture attachments, sixteen retained staples through 32 leg attachments, 32 cured collar-sector attachments, conserved PBD leak particles, patency evaluation, an eight-second pressure-decay challenge, at least 120 CUDA steps, and finite state. See docs/adaptive_anastomosis_robot/VALIDATION.md.

Adaptive seal-and-divide robot

The seal-and-divide system integrates tissue centering, symmetric jaw compression, guarded blade travel, irrigation, suction, energy-state estimation, thermal and impedance observables, seal verification, and explicit blade-before-seal interlocks.

The native CUDA evidence gate requires two cooked vessel surfaces, two distal fixtures, sixteen bridge attachments, four temporary compression attachments, four retained seal-band attachments, interlocked division, release of temporary constraints, exact joint counts, finite state for 120 steps, and zero engine errors in both standalone and Franka-mounted representations. See docs/adaptive_seal_divide_robot/VALIDATION.md.

SafePlane dissection robot

The SafePlane system combines bilateral distributed traction with blunt spreading, seven-port hydrodissection, guarded articulated micro-scissors, and a retractable low-energy spatula. Independent vessel, nerve, and duct assets retain explicit continuity and modality-specific clearance interlocks; an override produces inspectable simulated injury state rather than bypassing the physical model.

Its CUDA matrix covers the 17-DoF standalone mechanism and 24-DoF Franka assembly, two cooked tissue surfaces, target-bed fixtures, traction and bridge attachments, all 28 releasable adhesion bridges, protected-structure continuity, conserved PBD fluid, finite state for 120 steps, and zero engine errors. See docs/safeplane_dissection_robot/VALIDATION.md.

Perfusion and tissue-viability robot

The perfusion system registers stereo RGB, NIR/ICG, laser speckle, thermal, surface oxygenation, depth, Doppler, and ultrasound sensing around one TCP. Its estimator is blind to scenario labels and latent flow state, fuses temporal ICG evidence, removes failed modalities, tracks conserved contrast and coupling gel, and explicitly abstains on invalid registration, timing, coverage, or confidence.

The v0.1.1 CUDA matrix covers both the 12-DoF standalone mechanism and 19-DoF Franka assembly for 260 steps, six nonconstant rendered camera streams, finite depth, a cooked tissue surface with two fixtures, blind diagnosis of six modeled faults, force-coupled probe contact, evidence-based intervention, and three loaded-arm poses with the authored 2.537 kg payload. See docs/perfusion_viability_robot/VALIDATION.md.

Surgical-oncology training cell

The OncoSurgery Cell integrates a payload-backed 22-joint tumor-resection tool, a 3,028-cell liver tumor field, 96 explicit resection bonds, protected vascular and bile-duct interlocks, registered RGB/depth, NIR, hyperspectral, ultrasound, OCT, and Raman contracts, specimen containment and orientation, cavity verification, and corrective resection.

Its Isaac Lab runtime provides standalone, rigid-proxy, Franka-mounted, liver, specimen, and three-station workcell factories; bounded reset-time domain randomization; a 12-term policy observation; dense safety-aware reward; and a contract-gated final margin report. The imported USDA layers were repaired and wrapped with lightweight relative-payload interfaces, and the Franka payload representations now share a consistent authored 2.5534 kg mass. The native tissue route couples the task to the Dynamic Patient liver's explicit tetrahedral PhysX GPU volume deformable while retaining the registered resection graph for irreversible topology changes. Recorded native CUDA evidence now records a passing RTX 4090 non-contact volume-stability lane with 274 live tetrahedral nodes, bounded displacement and speed, and zero engine errors. Robot-tissue contact, rendered sensors, Franka payload behavior, and all real-world, biomechanical, and clinical evidence remain explicit promotion gates. See docs/SURGICAL_ONCOLOGY.md.

Platform architecture

Clinician / researcher
        │
        ▼
Doctor Studio ── control, guidance, study configuration, review
        │
        ▼
Dr.Anmar hub ── authentication, operator lease, lifecycle, provenance
        │
        ▼
Isaac worker ── task, robot, sensor, controller, recorder bindings
        │
        ├── Isaac Lab articulation and task APIs
        ├── PhysX rigid, deformable, attachment, and particle mechanics
        ├── OpenUSD scenes, materials, assets, and variants
        └── optional bounded NVIDIA / SonoGym workflows
        │
        ▼
Evidence ── trajectories, manifests, metrics, logs, dataset cards

This separation is intentional:

• the browser never determines physical contact, attachment, puncture, division, or task success;
• visible controls are converted into bounded robot commands and audited;
• native simulator state is the authority for mechanics and outcomes;
• provider-specific behavior remains behind explicit adapters;
• downloaded assets, demonstrations, checkpoints, logs, and runtime state remain outside Git; and
• promotion claims are limited to the evidence recorded for the exact tested configuration.

Doctor Studio

Doctor Studio presents the simulator as a procedural workspace rather than an infrastructure console. It includes:

• guided robotics lessons expressed in clinical language;
• live OpenUSD operating and simulation-training rooms;
• keyboard and game-controller bimanual control;
• camera-native one- or two-hand webcam teleoperation;
• bounded voice commands with a matching typed-command fallback;
• immediate stop, pause, takeover, and camera controls;
• demonstration recording, replay, and clinician-selected references;
• Skills Twin trajectory and phase analysis;
• seeded Failure Lab perturbations and policy evaluation; and
• multimodal study manifests for RGB, depth, segmentation, point clouds, wrist cameras, pose, torque, contact, deformation, operator input, and procedure annotations.

The interaction model and safety behavior are documented in docs/KEYBOARD_CONTROLS.md and docs/WEBCAM_TELEOPERATION.md. The multimodal data contract is described in docs/MULTIMODAL_STUDIES.md.

Evidence model

Dr.Anmar uses five distinct evidence levels:

Level	Establishes	Does not establish
Product capability	The asset or workcell is integrated and available for its stated simulation-training workflow	Numerical fidelity
Repository verification	Assets, schemas, manifests, hashes, paths, controller invariants, and package consistency pass	Native engine behavior
Native-simulator evidence	A named revision ran on a recorded Isaac/PhysX stack and hardware configuration	Generalization or real-world behavior
Real-world evidence	Instrumented hardware, materials, sensors, or wet/dry-bench measurements support a specific correlation claim	Clinical effectiveness
Clinical evidence	A defined clinical study and review support a specific clinical claim	Claims outside that study

Dr.Anmar’s current robot evidence is repository and native-simulator evidence. Mechanical constants, tissue parameters, pressure/flow thresholds, energy models, contact limits, damage proxies, and success thresholds remain engineering parameters unless a robot-specific artifact records instrumented real-world evidence.

Requirements

The simulator runtime requires a Linux x86-64 system with a compatible NVIDIA GPU. Source, documentation, and browser code can be inspected on macOS or Windows, but this project's Isaac Sim backend does not execute there.

Recorded native-simulator lanes currently include:

• Isaac Sim 5.1 with Isaac Lab 2.3.2 for the stable operating-room workflow; and
• Isaac Sim 6.0.1.0 with Isaac Lab 6.1.16 for the surface-deformable robot evidence lane.

Python 3.10 or newer is required inside the corresponding Isaac environment. NVIDIA components and optional provider assets retain their own licenses and are not redistributed by this repository.

Quick start

Clone the repository outside the Isaac Lab checkout:

git clone https://github.com/Numi2/drAnmar.git
cd drAnmar
cp .env.example .env

Set ISAAC_PYTHON in .env to the Python executable in the selected Isaac environment. Runtime data defaults to ~/.local/share/dr-anmar; change DR_ANMAR_ROOT to relocate it. For a shared workstation, configure a long random DR_ANMAR_ACCESS_TOKEN, use DR_ANMAR_COOKIE_SECURE=1 behind HTTPS, and keep the services on a trusted LAN or private VPN.

Install the local extensions:

export IsaacLab_PATH=/absolute/path/to/IsaacLab
./orbitsurgical.sh

Start Doctor Studio:

./dr_anmar_suite.sh start

Open http://localhost:2360. Service controls are:

./dr_anmar_suite.sh status
./dr_anmar_suite.sh logs
./dr_anmar_suite.sh restart
./dr_anmar_suite.sh stop

See SECURITY.md before allowing access from another machine.

Reproducing native-simulator evidence

Run a robot's static validator and unit tests before its CUDA runtime test. For example:

python3 scripts/validate_dranmar_wound_preparation_robot.py --require-usdchecker
python3 -m unittest -v tests/test_wound_preparation_robot.py

./isaaclab.sh -p examples/validate_wound_preparation_runtime.py \
  --headless --device cuda:0 --representation standalone
./isaaclab.sh -p examples/validate_wound_preparation_runtime.py \
  --headless --device cuda:0 --representation franka

Equivalent validators and runtime programs are included for the exposure, hemostasis, anastomosis, seal-and-divide, SafePlane dissection, and perfusion viability systems. Do not transfer a passing result between robot revisions, representations, simulator versions, GPUs, or physics configurations.

The Isaac Lab documentation GIFs can be regenerated with:

./isaaclab.sh -p scripts/capture_dranmar_robot_gif.py \
  --headless --enable_cameras --device cuda:0 \
  --robot adaptive-hemostasis \
  --output docs/screenshots/robots/adaptive-hemostasis-isaac-lab.gif

Valid robot identifiers are wound-preparation, atraumatic-exposure, adaptive-hemostasis, adaptive-anastomosis, and adaptive-seal-divide.

For a provider asset already mirrored onto the runtime machine, pass --franka-usd /absolute/path/to/franka.usd; omit it to use Isaac Lab's configured Franka asset location.

Command-line workflows

# Inspect registered environments and curriculum content
./dr_anmar.sh list
./dr_anmar.sh catalog
./dr_anmar.sh doctor

# Run a bounded task smoke session
./dr_anmar.sh smoke Isaac-Lift-Needle-PSM-IK-Rel-v0 120

# Start a training experiment
./dr_anmar_train.sh rsl_rl \
  Isaac-Lift-Needle-PSM-IK-Rel-v0 \
  --num_envs 256 \
  --max_iterations 1000

Repository structure

web/                  Doctor Studio browser application
scripts/              Hub, workers, control adapters, generators, and checks
examples/             Native CUDA evidence programs
tests/                Controller and package regression tests
source/extensions/    Simulator tasks and articulated robot assets
source/standalone/    Teleoperation, data, training, and policy workflows
physics_next/         Versioned next-generation physics contracts
docs/                 Architecture, mechanisms, evidence, and validation records
dr_anmar_*.sh         Portable service, runtime, and training launchers

Development checks

Run the public-release checks before submitting changes:

python3 scripts/check_public_release.py
python3 scripts/audit_project_consistency.py
python3 scripts/audit_keyboard_controls.py
python3 scripts/check_web_syntax.py
python3 -m compileall -q scripts source
bash -n dr_anmar.sh dr_anmar_suite.sh dr_anmar_train.sh \
  dr_anmar_workstation.sh orbitsurgical.sh

See CONTRIBUTING.md, NOTICE.md, and the validation backlog.

Citation

If ORBIT-Surgical-derived components contribute to published research, cite:

@article{yu2024orbit,
  title={ORBIT-Surgical: An Open-Simulation Framework for Learning Surgical Augmented Dexterity},
  author={Yu, Qinxi and Moghani, Masoud and Dharmarajan, Karthik and Schorp, Vincent and
          Panitch, William Chung-Ho and Liu, Jingzhou and Hari, Kush and Huang, Huang and
          Mittal, Mayank and Goldberg, Ken and others},
  journal={arXiv preprint arXiv:2404.16027},
  year={2024}
}

Publications using Dr.Anmar should additionally report the repository revision, robot asset manifest, simulator and Isaac Lab versions, GPU/driver, scenario and seed, control policy, sensor profile, and applicable evidence artifact.

License

Dr.Anmar and the included ORBIT-Surgical-derived source are distributed under the BSD 3-Clause License. Isaac Sim, Isaac Lab, NVIDIA assets, SonoGym, and other optional dependencies retain their own licenses and terms.