We analyze sensor data and motion logs from robot platforms to improve autonomous behavior and task execution precision.

• Sensor-fusion-based real-time state estimation
• Environment-adaptive motion planning
• Autonomous behavior under safety constraints

We learn from robot manipulation data to enable autonomous complex task execution and improve operational efficiency.

• Demonstration-based manipulation learning pipeline
• General-purpose manipulation across diverse tasks
• Real-time obstacle-avoidance path planning

We use manipulation data collected from teleoperation systems to accelerate robot learning and transfer across tasks.

• Teleoperation-data-based behavior cloning
• Bimanual coordinated task learning
• Sim-to-real transfer learning

We fine-tune pretrained VLA foundation models on domain-specific demonstration data to adapt them to new tasks and sites quickly.

• Adapter/LoRA fine-tuning on open VLA models (OpenVLA, π0, RT-2)
• Aligned data pipeline for site cameras, language instructions, and trajectories
• Domain adaptation for zero- and few-shot generalization to new tasks

Instead of separating perception, planning, and control, we learn a single policy from sensor input to motor commands, removing error accumulation between modules.

• Unified policies over image, proprioception, and language inputs
• Diffusion-policy and transformer-based action generation
• Distilled, lightweight pipelines for on-device inference