Research map / Reinforcement learning / Imitation learning
Instant Policy: In-Context Imitation Learning via Graph Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Instant Policy, a graph-diffusion model for in-context imitation learning that learns new manipulation tasks from one or two demonstrations without weight updates, trained purely on simulated pseudo-demonstrations.
Problem
In-context imitation learning for robot manipulation: enabling a robot to perform a new task immediately from one or two demonstrations at test time, without retraining, while overcoming the lack of large robotics datasets and poor inductive biases for 3D action representation.
Approach
Represent demonstrations, current segmented point-cloud observation and future gripper actions as a heterogeneous graph, cast action prediction as conditional graph diffusion over that graph, and train on procedurally generated pseudo-demonstrations (random object-centric waypoint trajectories in simulation) as an effectively unlimited training source.
Key results
- On 24 RLBench tasks, Instant Policy averages 0.97 success on seen tasks (PD++) and 0.82 on unseen tasks, versus at most 0.65 and 0.43 for baselines.
- Real-world average success over 16 tasks is 88.75% versus 34-41% for BC-Z*, Vid2Robot* and GPT2*.
- Generalisation to novel object geometries rises from 40% with one demo to 90% with four demos in the context.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Instant Policy instead of the strongest reported baseline | RLBench, PD only · RLBench, all tasks | success rate: 0.38 -> 0.71 | Helped |
| Use Instant Policy instead of the strongest reported baseline | RLBench, PD++ · RLBench, seen tasks | success rate: 0.65 -> 0.97 | Helped |
| Use Instant Policy instead of the strongest reported baseline | RLBench, PD++ · RLBench, unseen tasks | success rate: 0.43 -> 0.82 | Helped |
| Use Instant Policy instead of the strongest reported baseline | Real-world tasks · 16-task average | success rate: 40.63 -> 88.75 | Helped |
| Increase the number of demonstrations across different object geometries | Instant Policy · Novel object geometries | average success rate: 40 -> 90 | Helped |
| Use decoupled translation and rotation action representation (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -15 -> 0 | Helped |
| Use decoupled translation and rotation instead of translation and quaternion (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -37 -> 0 | Helped |
| Use decoupled translation and rotation instead of translation and angle-axis (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -21 -> 0 | Helped |
| Predict flow rather than added noise (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -7 -> 0 | Helped |
| Use diffusion rather than direct action regression (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -29 -> 0 | Helped |
| Predict multiple future actions rather than one (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -52 -> 0 | Helped |
| Use more than one diffusion step (ablation) | Instant Policy · RLBench, 24 unseen tasks | success-rate change: -16 -> 0 | Helped |
Surprising findings
- Models trained only on pseudo-demonstrations already achieve non-zero success on all methods, with Instant Policy averaging 0.71 on 24 RLBench tasks without any real demonstrations.
- Performance does not improve beyond two demonstrations when test objects match training geometry, and a single-network variant or fully connected graph performed significantly worse.
Limitations
- Assumes segmented point clouds and relies on Markovian short-horizon tasks; no collision avoidance or full configuration-space control.
- Sensitive to demonstration quality and downsampling at inference, and lacks precision for low-tolerance or contact-rich tasks.
- Modality transfer to language yields behaviour sometimes mismatched with object geometry in the environment.
Experimental setup
Datasets: ShapeNet RLBench pseudo-demonstrations (simulated) real-world kinesthetic demonstrations for co-fine-tuning
Benchmarks: RLBench (24 tasks) real-world Sawyer everyday tasks (16 tasks) novel object generalisation (4 tasks)
Metrics: success rate validation loss
Baselines: BC-Z Vid2Robot GPT2-style model
Builds on: DDIM Set Abstraction (PointNet++) heterogeneous graph transformer occupancy networks Perceiver Resampler
Closest papers by meaning
- Policy Decorator: Model-Agnostic Online Refinement for Large Policy Model (ICLR 2025 · similarity 0.89)
- Deep Demonstration Tracing: Learning Generalizable Imitator Policy for Runtime Imitation from a Single Demonstration (ICML 2024 · similarity 0.88)
- Learning with Language-Guided State Abstractions (ICLR 2024 · similarity 0.88)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.88)
- Text-Aware Diffusion for Policy Learning (NeurIPS 2024 · similarity 0.88)
- Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.