← Back to the blog
July 2026

Robot Foundation Models in the factory: Pi-Zero, OpenVLA, GR00T N1.5 — the end of teach-pendant programming

For 50 years an industrial robot learned a single task through teach-pendant trajectory programming. In 2024–2026 three open foundation models for robotic manipulation arrived — Pi-Zero from Physical Intelligence, OpenVLA from Stanford and GR00T N1.5 from NVIDIA — promising to end that era. The operator shows the robot what to do in VR/AR through 5–10 demonstrations, and a generic model translates that into a real-time trajectory. This article walks through how it actually works today, where it is enough, where it is not, and the real implications for MES integrators across the EU.

📅 July 6, 2026🏭 AI✒ Martin Szerment
Robot Foundation Models in the factory: Pi-Zero, OpenVLA, GR00T N1.5 — the end of teach-pendant programming

For 50 years an industrial robot learned one task through teach-pendant trajectory programming — point by point, angle by angle, hours of integrator work for a single pick-and-place operation. A second product variant meant a second program. A third — a third one. That was the reality of FANUC, KUKA, ABB and Yaskawa from the 1980s onwards.

In 2024 and 2025 three projects rewired that reality. Pi-Zero (Physical Intelligence, November 2024) — the first open foundation model for robotic manipulation, trained on 10,000 hours of demonstrations across many robot platforms. OpenVLA (Stanford, June 2024) — a 7-billion-parameter Vision-Language-Action model that fuses an LLM with a motor policy. GR00T N1.5 (NVIDIA, Computex 2026) — the first production-grade NVIDIA foundation model for humanoid and manipulator robots. All three with open weights.

In practice, the integration model for robots on a production line changes. The operator shows the robot in VR what to do (teleoperation, 5–10 demonstrations), the model translates it into a real-time trajectory, the MES stores the task as a prompt plus few-shot demonstrations. A new product variant on the line? The robot pulls the prompt from MES and generates motion with no programmer intervention.

Below: each of the three models, where it works today, where it does not, and what it means for European MES integrators.

What it looks like today — the teach pendant in 2026

The standard procedure for bringing up a new robotic cell in a European plant in 2026 looks roughly like this:

  1. The integrator spends 3–5 days on the teach pendant programming trajectories — pickup positions, transits, drop-offs — for each variant of the part (typically 5–30 variants per line)
  2. Quality validation — another 2–3 days of testing on the real material, debugging collisions, finding the sweet spot of speed vs grip quality
  3. Documentation — the robot code (KAREL for FANUC, KRL for KUKA, RAPID for ABB) goes into a repository, plus a variant-change instruction for the operator
  4. Variant change in production — the operator picks a program from the teach-pendant menu or triggers it via MES — the variant change takes 30–90 seconds (typically program reload plus parameters)

Cost of one cell: EUR 3,500–10,000 for the robot plus EUR 5,000–15,000 for integrator work — depending on complexity. Deployment time 4–8 weeks.

The weaknesses of this model have been known for decades:

  • No generalization. A robot taught to pack 0.5 L bottles cannot pack 0.7 L bottles without a new program, even if the difference is minimal
  • No adaptation to process variance. Part slightly misaligned on input? The trajectory does not fit, collision, production stop
  • Cost of change. Introducing a new SKU = a fresh 4–8 week integrator project. That disqualifies robotization for low-volume plants (under 5,000 units of one variant)
  • Vendor lock-in. A KAREL program does not run on a KUKA. A customer with a FANUC cell pays 2× more to switch suppliers

Pi-Zero — the first generic robot policy

Physical Intelligence (founded 2024 in San Francisco, USD 400M Series A) released Pi-Zero in November 2024. It is a foundation model for manipulation — analogous to GPT for text — trained on 10,000 hours of demonstrations across multiple robots: Universal Robots cobots, ALOHA (dual-arm system), Franka Panda, some humanoids.

Architecture: a VLM (Vision-Language Model) backbone plus a flow-matching action head for trajectory generation. Input: RGB camera plus a natural-language prompt ("pick up the red cup and place it in the box"). Output: a sequence of actions in joint space at 50 Hz.

What matters for MES: few-shot transfer. A model pretrained on 10k hours of generic data can learn a new task from 5–10 demonstrations (teleoperation via a VR controller). No programming, no teach pendant. Success in Physical Intelligence demos: shirt folding (a classical textile task long considered very hard), multi-variant packing, tool handing.

Pi-Zero received open weights in July 2025 (HuggingFace pi0). Reference PyTorch implementation, integration with Hugging Face's LeRobot library. Entry barriers for an integrator are low — a GPU with 24 GB of VRAM (RTX 4090 or Jetson AGX Thor) is enough for inference; fine-tuning needs an A100/H100 cluster with 4–8 nodes.

Real case: Covariant (acquired by Amazon in 2024) uses a Pi-Zero derivative in warehouse picking — 95% recall in consumer-goods packing, with 12-hour fine-tuning for a new assortment instead of 2-week integration.

OpenVLA — academic baseline, but it works

OpenVLA from Stanford and Berkeley is a 7B VLA (Vision-Language-Action) model — Llama-2-7B as the backbone + ViT (Vision Transformer) as the image encoder + a linear action head. Trained on Open X-Embodiment — an aggregated dataset from 21 academic institutions, 970,000 trajectories across 22 robot platforms.

What is interesting for industry: total openness. Code, training data, weights, fine-tune scripts — all on GitHub under the MIT licence. You can take it and stand it up internally without legal back-and-forth.

In practice OpenVLA is worse than Pi-Zero on multi-step manipulation, but sufficient for one-shot pick-and-place (grip + transfer + drop). Prediction latency 60–80 ms on RTX 4090, recall on typical industrial tasks 80–88%.

MES use case: operator-assist in training mode. The operator shows the robot a new part variant through 3–5 VR demonstrations, OpenVLA fine-tunes for 30 minutes, the result is enough for low-stakes production (cosmetics, packaging).

GR00T N1.5 — NVIDIA at Computex 2026

GR00T N1.5, unveiled at Computex 2026 (covered in our keynote retrospective), is the production-grade version of NVIDIA's foundation model. What changed from N1 (March 2026):

  • 300% more training data, a substantial portion of which is synthetic generated in Cosmos (NVIDIA's world model)
  • Zero-shot manipulation across 50+ gripper types — from two-finger to suction to multi-finger
  • 80 ms latency on Jetson Thor (was 220 ms on Jetson Orin AGX for N1)
  • TwinCAT and Beckhoff support in the default deployment kit — critical for European industry, where these platforms dominate the automation side

GR00T N1.5 has the strongest partnerships of the three models. NVIDIA confirmed cooperation with FANUC, KUKA, Boston Dynamics, Agility Robotics. That means in 2027 you will likely buy a FANUC cobot with GR00T as the default integration option, not classical teach-pendant flow.

Trade-off: GR00T is open weight, but training and fine-tuning require NVIDIA hardware (Jetson Thor or a server with H100/B200). For smaller plants that means an additional USD 5–20k on hardware vs OpenVLA on a consumer GPU.

What actually changes for an MES integrator

Four concrete changes in 2026–2027:

1. Deployment time drops from weeks to days. Instead of 4–8 weeks of trajectory programming — 2–5 days: hardware setup, 5–10 demonstrations per task, fine-tune (30 min to 4 hours), validation. For a plant with 5–10 robotic cells per year, that is a saving of roughly EUR 50–100k per year on integrator work.

2. Variant change = new prompt, not new project. The operator shows the robot in VR a new variant through 5–10 demonstrations. The robot generalizes, MES stores the prompt as a task parameter. Time: 30–60 minutes instead of weeks. That eliminates the "low-volume isn't worth robotizing" argument — the breakeven point falls from 5,000 units to roughly 500 units of a single variant.

3. MES becomes a store of prompts. A new MES responsibility: maintaining a library of prompts per product variant, versioning them, validating through approval workflow (who can modify a prompt to a robot? — a new RBAC tier). This fits the existing OmniMES stack — a prompt is just text plus references to VR-recorded demonstrations.

4. Vendor lock-in drops. Pi-Zero, OpenVLA, GR00T — all three open weight, all support multiple robot manufacturers. Switching from FANUC to KUKA with a fine-tuned foundation model takes 4–8 hours of re-training rather than a full rewrite in another DSL.

Where it does NOT work — an honest list

Four cases in which the teach pendant still wins in 2026:

1. High-precision (under 0.1 mm tolerance). Foundation models handle manipulation on a 1–10 mm scale. For precise electronic assembly (SMT, micro-soldering, fibre-optic alignment) their accuracy is too low. Here classical programming with force-torque feedback and accurate calibration wins by an order of magnitude.

2. Cycle time under 2 seconds per pick. Prediction latency of 80 ms (GR00T) plus planning plus execution is typically 0.8–1.5 s per operation. For high-throughput packing (e.g. a bottling line at 12,000 bottles/h) classical PLC-driven trajectories run at 200–400 ms — 3–5× faster.

3. Safety-critical in direct human collaboration. Foundation models do not yet hold ISO 10218 / ISO 13849 certification for safety-rated operation. A standard cobot with a certified safety stop operates under ISO. A foundation model as a control loop for a cobot needs a separate safety layer built from scratch.

4. Very high payload (over 50 kg). Models are mostly trained on cobots and light industrial arms (under 20 kg payload). 100+ kg palletizing, automotive body-in-white welding — here a classical industrial robot with offline planning in RobotStudio/RoboGuide is still the right choice.

Reference architecture: MES + Robot Foundation Model

A stack that ties a foundation model to MES in a sensible way:

PLC / Robot Controller (FANUC R-30iB, KUKA KR C5, ABB OmniCore)
  ↕ EtherCAT / PROFINET (real-time, unchanged)
GR00T-compatible Hardware (Jetson Thor or edge GPU server)
  ↕ ROS 2 / DDS bridge
Foundation Model Inference (Pi-Zero / OpenVLA / GR00T N1.5)
  ↕ MQTT / Kafka events
MES Layer (OmniMES)
  ├── Prompt Store (per product variant, in PostgreSQL)
  ├── Demonstration Library (VR recordings, S3-compatible)
  ├── Fine-tune Pipeline (orchestrated jobs)
  ├── Approval Workflow (RBAC for prompts)
  └── Performance Monitoring (success rate per prompt, drift detection)

The key element is MES as a store of prompts and demonstrations. Each robotic task is:

  • A language prompt (e.g. "pick the red package from the conveyor and place it in slot 3")
  • 5–20 VR demonstrations per prompt (reference for retraining)
  • Fine-tune result as an artifact (model weights, ~200 MB)
  • Production telemetry (success rate, cycle time, collisions)

That maps cleanly onto OmniMES data architecture — prompts in PostgreSQL, demonstrations in object storage, telemetry in TimescaleDB as time-series.

What it means for European integrators

FANUC, KUKA, ABB, Universal Robots — all four announced foundation-model integrations in 2026–2027. For regional European integrators it is two paths:

Path A: stay a classical integrator. Teach-pendant programming for high-precision / high-throughput / safety-critical remains. Margins on the work do not disappear — they actually grow, because the market has fewer people willing to write KAREL/KRL/RAPID for genuinely hard cases. Specialization as the strategy.

Path B: learn foundation models, offer hybrid integration. The customer gets a classical robot with a foundation model as the default flow plus a classical fallback. Marketing margins are high — it is a new generation of offering with little regional competition through end-2027. Requires investment in skills (PyTorch, ROS 2, NVIDIA stack) and demo hardware (Jetson Thor + a cobot for POC).

A realistic forecast: by end-2027 around 30% of new robot deployments in European mid-market will have a foundation model as default, 70% will remain classical. By 2030 those proportions reverse.

Takeaways for the production director

Three things to remember:

First, foundation models in robotics in 2026 are not hype — this is a mature technology with three production open-weight models (Pi-Zero, OpenVLA, GR00T N1.5) and partnerships with large robot manufacturers (FANUC, KUKA, ABB). Realistic deployment path: 12–18 months from decision to first production POC.

Second, the strongest change is the economics of low-volume. Foundation models make robotization of products that were previously manual (under 5,000 units of a variant) economically viable. For European custom manufacturing (niche automotive, consumer electronics, premium packaging) — a fundamental shift.

Third, MES becomes part of the robotic stack, not just an observer. New responsibilities: store of prompts, demonstration library, approval workflow, performance monitoring. That requires extending the existing MES or replacing it with an AI-ready one. A TimescaleDB-based data architecture helps — telemetry from robots is classical time-series that hypertables handle unmodified.

Foundation models will not replace classical robot programming in 2026. But by 2028 they become the default integration option for most production deployments. European mid-market manufacturers that start experimenting in 2026 will have a two-year edge over regional competitors.

Sources