A voice-directed bimanual robot-arm system that plans with an LLM, acts in simulation, verifies its own results with a real Intel OpenVINO vision model — and when it fails, says so honestly and fixes itself.
We built one that shows what happens when it doesn't.
Physical-AI demos are everywhere right now, and bimanual manipulation is the obvious "hot" category. That means a crowded field of lookalikes: "say a command, watch an arm move." Most of them optimize for exactly one thing — looking like it worked.
"In real factories and warehouses, the expensive failures aren't the obvious ones — they're the silent ones a system doesn't know it made."
EchoGrasp optimizes for the inversion: not "look, it worked" — but "look, it knows when it didn't, and fixes it."
Three real infrastructure services — speech, planning, vision — wired through an orchestrator that has to handle partial failure gracefully.
Step 1's grasp genuinely misses. The system says so in plain language, re-plans, retries, and succeeds. Every confidence number on screen is a real OpenVINO INT8 inference result, and the pass/fail call is confirmed against the simulator's own physics state.
Captured live from the running dashboard: staged failure on step 1 → plain-language reason → corrected retry → successful handoff → successful placement → 100% session score.
We'd rather tell you the real accuracy of our own verification model than oversell it — that felt like the point of the whole project.
This page is a static project showcase — GitHub Pages can't run Python, physics simulation, or AI inference. The actual app is a real local server. It takes about two minutes to get running.
git clone https://github.com/JoysonJoe20/InfraSummit.git cd InfraSummit python -m venv .venv && source .venv/bin/activate pip install -r requirements.txt cp .env.example .env # add your SPEECHMATICS_API_KEY and GROQ_API_KEY python -m backend.verify.train_or_prepare_model # one-time: trains the OpenVINO verifier python -m backend.verify.export_openvino ./scripts/run_demo.sh # opens http://localhost:8000