Simulation & transfer
Domain adaptation
Domain adaptation is a form of transfer learning that learns to perform a task in a target domain using knowledge from a related source domain whose data distribution differs. The task is typically shared or closely related, while observations, dynamics, environments, sensors, or other domain conditions change.
Also known as: domain-adaptive learning, robot domain adaptation
Updated
The domain changes while the task remains related
A source domain and target domain can differ in camera appearance, lighting, object distribution, sensor noise, dynamics or robot embodiment. Ganin and colleagues present a method that learns features useful for the source task while making source and target examples difficult to distinguish.
In robotics, common cases include synthetic-to-real perception, one laboratory to another, or one camera configuration to another. Adaptation can use labelled or unlabelled target data; the amount and type of target supervision should always be stated.
Adaptation differs from randomisation and generalisation
Domain randomisation varies training conditions in the hope that the target lies within the learned range. Domain adaptation explicitly uses information about the target domain during learning or adjustment. Domain generalisation instead aims to perform on unseen target domains without adapting to their data.
The boundaries can blur in a pipeline that randomises simulation, collects unlabelled real observations and then fine-tunes. Describe the actual data flow rather than using the terms interchangeably.
Alignment can erase useful distinctions
Matching aggregate feature distributions does not prove that corresponding objects, states or actions have been aligned. A representation can appear domain-invariant while losing task-relevant information, and shifts in robot dynamics cannot always be solved by changing visual features.
Evaluate on held-out target tasks and conditions, with target-only and no-adaptation baselines. Document which target data influenced training, calibration or model selection. Otherwise a claimed zero-shot result may include hidden target supervision.
Sources
Related terms
Models & learning
Transfer learning
Transfer learning uses knowledge learned from one or more source tasks or domains to improve learning or performance on a different target task or domain. In robotics, transfer may reuse representations, model weights, skills, data, or policies across robots, tasks, environments, sensing configurations, or simulation and reality.
Simulation & transfer
Sim-to-real
Sim-to-real is the transfer of a model, policy or behaviour developed wholly or partly in simulation to a physical robot or real environment. The central problem is the reality gap: errors in simulated dynamics, sensing, appearance and timing can make a strategy successful in simulation but unreliable or unsafe on hardware.
Simulation & transfer
Domain randomisation
Domain randomisation is a simulation-training technique that varies selected visual, physical or sensor parameters across generated examples or episodes. By exposing a model or policy to a distribution of environments rather than one calibrated scene, it aims to make real conditions fall within the learned variation and improve sim-to-real transfer.
Models & learning
Robot generalization
Robot generalization is a policy's ability to perform a learned behavior when test conditions differ from training, such as with new objects, environments, tasks, or robot embodiments. A useful claim specifies which conditions were held out.
Data & collection
Cross-embodiment data
Cross-embodiment data is robot training data drawn from multiple physical embodiments, such as arms, mobile manipulators, quadrupeds or humanoids with different kinematics, sensors and action spaces. The datasets are aligned or packaged so models can learn jointly from experience produced by different robots, although shared formatting does not make their observations or controls physically equivalent.