Robotics

Instant Policy

2024ResearchPublished: 28 October 2026Updated: 28 October 2026Published

Key innovation

Treats one or two demonstrations as in-context examples and generates robot actions through a learned graph-diffusion process, letting a robot acquire a new task instantly, without any per-task retraining.

How it works

1) Graph representation: from segmented point-cloud observations a heterogeneous graph is built that jointly expresses the context (demonstrations), the current observation and future actions, introducing suitable inductive biases and enabling structured reasoning over robot–environment relationships. 2) ICIL as graph generation via diffusion: a distribution over graphs is learned with a diffusion process using forward/backward Markov chains — during training noise is added to the representation and the model learns to remove it. 3) Inference: starting from actions sampled from a normal distribution, the model iteratively updates (denoises) only the parts of the graph that represent robot actions, implicitly modelling the conditional action distribution (e.g. 8 future actions). 4) Data: trained on pseudo-demonstrations — arbitrary trajectories generated in simulation — as a virtually infinite data pool.

Problem solved

Classic imitation learning / behavioural cloning encodes each task into network weights and needs many demonstrations plus costly per-task training. Instant Policy removes per-task retraining: new skills are acquired instantly from 1–2 demonstrations supplied as context, drastically cutting the time to deploy a new robot behaviour.

Components

Heterogeneous graph representationRepresentation binding demonstrations, observation and actions into one structure

A graph built from segmented point-cloud observations that jointly represents the context (demonstrations), the current observation and future actions, introducing inductive biases for structured reasoning.

Learned graph diffusion processAction generator (denoising) conditioned on context and observation

A diffusion model learning a distribution over graphs (forward/backward Markov chain). At inference it iteratively denoises only the action nodes, implicitly modelling the conditional distribution of robot actions.

Simulated pseudo-demonstrationsTraining-data source

Arbitrary trajectories generated in simulation, used as a virtually infinite pool of training data instead of hand-collected per-task demonstrations.