Models increasingly connect language and vision to bounded physical actions.
Robotics future signal · Signal 01
From scripted automation to learned skills
What cross-embodiment policies and language-conditioned robot models change—and why reusable skills still require task-level evidence on each machine and site.
Evidence statusActive research with selected multi-robot demonstrationsLast reviewed: 11 August 2026
01
Current FUURAA judgement
The direction is credible, but transfer is a measured capability—not a default property of a foundation model.
Unpack the forecast
Separate what has happened, what may happen and what can test the judgement.
Reusable skills across robots and sites could reduce reprogramming cost.
Transfer success, recovery rate and human intervention per operating hour.
Evidence map
Four evidence layers explain why this signal is worth tracking.
Each layer states what is observed and how far that observation can travel.
Cross-embodiment transfer
Open X-Embodiment and later first-party systems report policies trained across different robot types and tasks.
Language as task interface
Vision-language-action systems can condition bounded action sequences on natural-language goals.
Embodiment still matters
Sensor layout, kinematics, action spaces, tools and contact physics differ across robots.
Recovery is the hard test
Long operation exposes disturbances and partial failures absent from curated demonstrations.
Verification checklist
Five checks make the signal continuously confirmable—or falsifiable.
Do not substitute launches, one-off demonstrations or isolated benchmarks for sustained operational evidence.
- 01
Hold out a robot and site
Evaluate on an embodiment, layout and object mix excluded from training and tuning.
- 02
Keep the task constant
Compare transferred and robot-specific policies on the same success, safety and cycle metrics.
- 03
Count all assistance
Include prompt edits, teleoperation, resets, scene preparation and recovery labor.
- 04
Test distribution shift
Vary object appearance, placement, lighting, friction, payload and instruction phrasing.
- 05
Measure learning cost
Report demonstrations, compute, validation and engineering time needed for the new site.
Scope boundaries
State what cannot be inferred from the current evidence.
- 01
A successful transfer among selected research robots is not zero-shot deployment on arbitrary hardware.
- 02
Language-conditioned action does not remove the need for independent physical constraints and protective functions.
- 03
First-party demonstrations do not establish long-duration reliability, economics or certification.
Sources and evidence status
Keep the original record, date, evidence identity and reading boundary visible.
Industry statistics, international standards, first-party research and commercial disclosures are not treated as the same kind of evidence.
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 · 2025-09-25
↗Gemini Robotics 1.5
Shows first-party examples of instruction following and motion transfer across selected embodiments.