PalmSim
Plantation Simulation Platform
Simulate. Train. Deploy.
GPU-accelerated digital twin of palm oil plantations — build realistic 3D environments, train robot policies via reinforcement learning, and deploy autonomous harvesting strategies with confidence.
End-to-End Simulation Stack
From photorealistic rendering to GPU-accelerated physics — everything you need to train, test, and validate autonomous harvesting robots before they touch a real tree.
Digital Twin Engine
Photorealistic 3D plantation environments built on NVIDIA Omniverse — every tree, slope, and drainage channel modeled from real-world data.
RL Policy Training
Train navigation and harvesting policies in Isaac Lab with thousands of parallel environments. Months of field experience in hours.
GPU-Accelerated Physics
Warp and Newton engines simulate soft-soil deformation, fruit-cutting mechanics, and 20-30 kg FFB drop trajectories at real-time speed.
Terrain Simulation
Muddy slopes, waterlogged paths, root protrusions, and uneven ground — the exact conditions that break real robots, replicated faithfully.
Weather & Climate
Tropical rainstorms, high humidity, morning fog, and sudden downpours. Test how weather degrades vision, traction, and gripper friction.
Lighting Conditions
Canopy-filtered dappled light at noon, overcast grey skies, golden dawn — each dramatically changes CV model performance.
Robot Modeling
Full kinematic and dynamic simulation of PalmBot: 20m articulated arm, cutting tool, gripper force, and mobile base on soft terrain.
Sim-to-Real Pipeline
Domain randomization and physics calibration ensure policies trained in simulation transfer reliably to physical robots in the field.
Test Every Condition Before Deployment
Palm oil plantations are among the harshest environments for autonomous robots. PalmSim lets you stress-test in simulation what would cost months and millions in the field.
Flat Terrain
Well-maintained roads, young palms, easy navigation baseline
Steep Slopes
Hillside plantations, 15-30° grade, traction challenges
Waterlogged
Post-rain standing water, soft mud, wheel sinkage risk
Dense Canopy
Mature palms 15-20m, heavy occlusion, limited GPS
Tropical Storm
Heavy rain, wind gusts, reduced visibility, wet surfaces
Real ↔ Sim Pipeline
A continuous loop: real-world data builds the simulation, simulation trains the robot, the robot collects better data. Each cycle makes both worlds more accurate.
Why Simulation-First Wins
One hour of GPU simulation generates more training data than a month of field trials. Domain randomization across terrain, weather, and lighting produces policies that generalize to conditions the robot has never physically encountered.
Navigation Policies
RL-trained path planning through mud, slopes, and dense canopy
Harvesting Strategies
Optimal cutting angle, grip force, and fruit-catch trajectory per tree
Synthetic Training Data
Millions of labeled images for CV models — no manual annotation needed
Technical Specifications
Built on NVIDIA's robotics simulation stack — the same technology powering autonomous vehicles, humanoid robots, and warehouse automation.
Digital Twin Layer
Photorealistic plantation environments reconstructed from satellite imagery and ground surveys. 50+ parametric palm tree variants, dynamic weather, and seasonal changes.
Training Layer
4,096+ parallel environments on a single GPU cluster. Train navigation, harvesting, and fleet coordination policies — achieving years of experience in days.
Deployment Layer
Validated policies export directly to PalmBot hardware. Continuous monitoring feeds real-world performance back into simulation for the next training cycle.
