Models & learning
In-context learning
In-context learning (ICL) is a model's ability to use examples or instructions supplied in its input context to guide its output without updating its model weights for that task. In robotics, the context can be a video demonstration that conditions a pretrained policy; it does not mean the robot learned the underlying skills from scratch from that clip.
Also known as: ICL
Updated
How it works
The model receives useful examples or task information as part of its input, then uses that context to produce an answer or action. In the language-model formulation described by Brown and colleagues, examples are included in the prompt rather than used to update the model through a new training run. The idea can be extended to other inputs: Skild AI describes S1 as using a video demonstration to condition a pretrained robot policy at inference time.
In-context learning for robots
A robot demonstration can show the task's order, relevant objects, contact points, and intended outcome more precisely than a short language instruction. The policy still needs prior capability from training; the video is a task-specific input that steers those capabilities. Skild's S1 research post reports this approach for manipulation tasks, while noting that the robot uses the same model weights during the examples shown.
The meaning of “unseen” depends on the evaluation. A task sequence may be new even when the model has previously encountered some of its individual actions. In-context learning claims should therefore describe what was held out: the task, objects, environment, embodiment, or some combination.
How it differs from fine-tuning
Fine-tuning changes model parameters using additional training examples. In-context learning supplies examples in the input for that run without making a task-specific weight update. A system can use both methods across its lifecycle: pretraining may teach broad skills, a prompt may specify the task at deployment, and later training may still update the model.
In-context learning is also related to, but not synonymous with, imitation learning. Imitation learning describes learning behavior from demonstrations; an in-context policy can use a demonstration at inference time, and its earlier pretraining may itself include imitation learning.
What to evaluate
For a robot system, ask whether performance was measured on genuinely held-out tasks and whether the demonstration came from a different scene or embodiment. Report full-task completion as well as per-step scores, count human interventions, and test recovery from errors. A single video prompt does not establish generality across robots, environments, or long-horizon tasks.
For background on reusable robot models, see robot foundation model. For the input-output rule that turns observations into actions, see policy.
Sources
Related terms
Models & learning
Robot foundation model
A robot foundation model is a broadly pretrained model intended to provide a reusable starting point for multiple robot tasks, environments or embodiments. It learns from diverse robotics and sometimes web or human data, then acts directly or is adapted with target-domain data. The term describes a training and reuse strategy, not one fixed architecture.
Models & learning
Imitation learning
Imitation learning is a family of methods that learns a policy from examples of expert behaviour rather than specifying every control rule by hand. In robotics, demonstrations pair observations or states with actions, trajectories or inferred objectives. Behaviour cloning is one imitation-learning method; interactive and inverse approaches address different supervision and distribution-shift problems.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.