See the interfaces among data, control, hardware and people.
Robotics technology stack · Layer 02
Embodied foundation models
A practical map of vision-language-action models, cross-embodiment data, robot policies and the gap between model demos and dependable work.
Evidence statusActive research; selective real-world demonstrationsLast reviewed: 11 August 2026
02
FUURAA thesis
Foundation models can broaden task priors and interfaces, but the deployed robot still needs bounded actions, real-time control, safety supervision and evidence on its own embodiment.
Reading method
Turn a technology label into an engineering chain, testable claims and explicit boundaries.
Translate capability into task, latency, failure and recovery.
Separate standards, independent measurement, research and first-party claims.
State what cannot be inferred from a demo, benchmark or interface.
System breakdown
Four interdependent layers determine whether the technology can enter real work.
Each layer shows its role and the failure signal most worth watching.
Multimodal representation
Images, language, robot state and action histories are encoded into a shared policy context.
Robot data
Demonstrations and trajectories teach contact-rich actions that web data cannot directly supply.
Policy and action interface
A model may output waypoints, skills or low-level actions, each with different latency and safety implications.
Runtime guardrails
Skill libraries, validators, monitors and human escalation bound what the model may do in the physical world.
Engineering evaluation
Five checks turn abstract capability into reviewable system evidence.
Record normal performance, failure, recovery and human cost—not only the best-looking result.
- 01
Separate model from system
Report model capability, controller, sensors, embodiment and operator assistance separately.
- 02
Hold out environments and objects
Evaluate on genuinely unseen combinations, not reordered scenes from the training distribution.
- 03
Count interventions and resets
Success rate should include retries, teleoperation, prompt changes and manual scene preparation.
- 04
Evaluate recovery
Introduce dropped objects, changed goals and partial failures rather than scoring only clean execution.
- 05
Constrain physical authority
Verify speed, force, workspace and tool constraints outside the generative model.
Scope boundaries
State what the evidence supports—and what it does not.
- 01
Cross-embodiment improvement in research trials does not imply zero-shot deployment on an arbitrary robot.
- 02
A first-party demonstration is not independent evidence of long-duration reliability or safety.
- 03
Language understanding can widen the attack and misuse surface as well as the task interface.
Sources and evidence status
Keep the source, date, evidence identity and reading boundary visible.
This page prioritises standards bodies, public measurement programmes, official project documentation and original research disclosures.
01
Google DeepMind · 2023-10-03
↗02Scaling up learning across many different robot types
Reports Open X-Embodiment data from 22 robot types and multi-lab RT-X evaluations.
Google DeepMind · 2025-03-12
↗Gemini Robotics brings AI into the physical world
Introduces Gemini Robotics and a reasoning-focused variant with selected demonstrations.