Learning-Based Robotics

Learning-Based Robotics is our service for developing, training, and validating robotic systems using data-driven and learning-based methods. In addition to classical imitation learning and reinforcement learning, we support the adaptation and fine-tuning of foundation models, including Vision-Language-Action (VLA) and other open-source models, to specific environments and applications.

Built on simulation-first development, this service covers the full learning lifecycle—from data generation and model adaptation to benchmarking and deployment on real robots—with a strong focus on robustness, reproducibility, and real-world applicability.

Data Generation & Collection

High-fidelity dataset generation (pxRobotLearning)
Supports large-scale dataset generation both in simulation and on physical robots, ensuring consistency between synthetic and real-world data distributions.
Teleoperation and human-in-the-loop data acquisition (pxTeleopForceXR)
Enables data collection through teleoperation interfaces, wearable devices, and interactive control, allowing humans to guide, correct, and intervene during task execution.
Multi-modal data support
Collects synchronized vision, point cloud, proprioceptive, force, and task-level signals for learning-based robotics.
Structured and versioned datasets
Provides standardized dataset formats with metadata, enabling reproducible training, benchmarking, and long-term evaluation.

Learning & Model Adaptation

Integrated IL and RL pipelines
Provides end-to-end imitation learning and reinforcement learning workflows, supporting both demonstration-driven and interaction-driven learning.
Fine-tuning of foundation models
Adapts pretrained foundation models to robotics-specific tasks, constraints, and sensor modalities.
Vision–Language–Action model adaptation
Fine-tunes VLA models to customer-specific environments, task semantics, and operational workflows.
Open-model integration
Supports integration and extension of open-source learning and perception models within a unified training framework.

Benchmarking & Validation

Simulation-based performance benchmarking
Evaluates learning performance under controlled, repeatable simulation scenarios.
Sim-to-real and sim-to-sim validation
Assesses policy robustness across different simulators and during transfer to real hardware.
Stress testing and edge-case evaluation
Validates behavior under disturbances, sensing noise, dynamic obstacles, and rare failure conditions.
Quantitative metrics and logging
Provides systematic evaluation metrics for policy stability, task success rate, and safety constraints.

Deployment & Inference

Model export and optimization
Supports model conversion, compression, and optimization for deployment on edge and embedded platforms.
ROS 2–based system integration
Seamlessly integrates trained models into ROS 2 pipelines for perception, planning, and control.
On-robot accelerated inference
Enables real-time inference on GPU- or accelerator-equipped robotic hardware with deterministic execution.

Typical Use Cases

Learning-based manipulation and navigation
Development of policies for grasping, manipulation, locomotion, and autonomous navigation.
Human–robot interaction and assistance
Training interactive behaviors that leverage human input, language instructions, and feedback.
Algorithm benchmarking and evaluation
Comparative evaluation of learning algorithms under standardized conditions.
Research and industrial R&D projects
Applied research, prototype development, and technology validation for industrial robotics applications (e.g., WPT project).