Research / Safety
How we think about safety and alignment
As logistics AI grows more capable, its decisions about people and vehicles carry more risk.
Our approach to safety is therefore tied closely to capability. We want to learn from systems as they develop, evaluate them against real operational conditions and increase safeguards as they move from analysis towards increasingly autonomous execution.
Progress in stages
We do not think autonomy in logistics arrives all at once.
There is a progression from systems that analyse operations, to systems that recommend actions, to systems that coordinate work and eventually interact with physical equipment and robotics.
Each stage creates different risks and requires a different level of testing, control and oversight.
Our research approach is iterative. We aim to learn from constrained deployments and real operational feedback before giving systems greater responsibility.
The impact of more capable systems
Greater intelligence can improve how logistics operations are planned, managed and executed.
It can also create greater consequences when a system is wrong.
We think about three broad areas of risk:
Incorrect decisions
A model may misunderstand an operation, draw the wrong conclusion or recommend an action that creates unintended consequences.
Loss of human control
As systems gain the ability to take actions rather than simply recommend them, people must remain able to understand, restrict and intervene in what those systems are doing.
Physical consequences
Errors become more significant when AI interacts with vehicles, equipment, robotics or other physical systems.
Our safety research therefore develops alongside the capabilities we are researching.
Our core principles
Embracing uncertainty
A system should know when it does not know.
Logistics environments are dynamic and incomplete information is common.
Our research explores how models can express uncertainty, identify missing evidence and defer to people rather than creating false confidence.
We treat safety as something that must be measured through testing and real-world evidence, not assumed from model behaviour in controlled conditions.
Defence in depth
No single safeguard should carry the entire burden of safety.
We favour multiple layers of control around increasingly capable systems.
These can include model-level safeguards, restricted permissions, operational rules, human approval, monitoring, auditability and fail-safe mechanisms.
As systems move closer to autonomous execution, those layers become more important.
Evaluation before responsibility
Greater capability should require stronger evidence.
We aim to evaluate models against realistic operational scenarios before increasing the level of responsibility they are given.
Our research considers areas such as:
- reasoning accuracy
- uncertainty calibration
- adherence to operational constraints
- robustness to incomplete or conflicting information
- intervention quality
- unexpected behaviour
- failure recovery
The standard required from a system recommending an action is different from the standard required from one executing it.
Methods that scale
Safety mechanisms need to remain useful as models become more capable.
Our research explores approaches that can scale alongside capability, including automated evaluation, continuous monitoring, structured permissions and systems that can identify when human intervention is required.
The aim is not simply to make today's systems safer, but to develop safety methods that remain effective as AI takes on more complex operational responsibilities.
Human control
Human authority remains central to how we think about AI in logistics.
AI should operate within clearly defined objectives, permissions and constraints.
People responsible for an operation should be able to understand what a system is doing, review consequential decisions and intervene when necessary.
Clear permissions
Systems should only be able to take actions they have explicitly been authorised to take.
Human escalation
When information is ambiguous, confidence is low or consequences are significant, systems should be able to escalate decisions to people.
Auditability
Important recommendations and actions should be traceable so operators can understand what happened and why.
Control in autonomous environments
As AI begins to interact with autonomous workflows, vehicles or robotics, operators must retain meaningful ways to monitor, restrict and stop those systems.
Learning from real operations
Safety cannot be understood entirely in a laboratory.
Real operations contain edge cases, incomplete information and interactions that are difficult to reproduce artificially.
We therefore see deployment and safety research as connected.
Carefully constrained real-world use can reveal where systems succeed, where they fail and where additional safeguards are required.
That feedback should inform the next generation of models, evaluations and controls.
Safety as capability grows
Our research with Depot-1 focuses on systems that understand operations and support operational decisions.
Looking further ahead, we are researching systems capable of executing and coordinating work across increasingly autonomous logistics environments.
The move from understanding to action changes the safety problem.
As systems gain more ability to affect the physical world, we believe the standards for reliability, oversight and control must rise with them.