What we’re actively working on.

Cross-embodiment VLA fine-tuning at the cell edge

OpenVLA matches RT-2 zero-shot on Open X-Embodiment, but customer-specific tasks need few-shot adaptation. We're benchmarking LoRA / DoRA fine-tunes against vendor-specific cobot SKUs (UR, ABB, FANUC) at <500 demonstrations per task.

Acoustic-emission fingerprinting for fleet monitoring

AE sensors detect bearing wear earlier than vibration but each asset has its own AE signature. We're studying few-shot enrolment protocols that reach 95% bearing-fault precision with <30 minutes of audio per asset.

Control-Informed RL for refinery loops

SAC and PPO are powerful but operators won't surrender PID. Control-Informed RL embeds PID set-point tracking inside the RL architecture so the controller gains nonlinear modelling capacity without losing engineering trust. We're applying this to catalytic reforming and gasoline blending against published 800,000 t/a baselines.

Vision-language QC with prompt-only adaptation

Qwen3-VL and Pixtral can adapt to new defect classes via prompt rather than retraining. We're measuring how this generalises — does prompt-only adaptation hold the 99%+ accuracy bar set by traditional supervised CNNs at production speeds, or only at lower throughputs?

Digital-twin-driven autonomous dispatch

Siemens Digital Twin Composer + NVIDIA Omniverse let us replay every dispatch in a photorealistic twin before commit. We're studying how much faster change-control cycles compress when the agent can validate against the twin without taking the line down.

Generative AI for IEC 61131-3 PLC ladder logic

Schneider and Siemens have shipped Copilots for PLC programming. We're researching reliability bounds — how often does generated ST/LD code introduce a deterministic bug that survives unit testing? What review patterns are needed before plant commissioning?