The Imitator Game logoThe Imitator Game

Benchmarking Robot Imitative Ability Beyond Action Prediction

Can robots imitate what a human intends — not just what they do? We introduce a four-level benchmark, a 20,000+ episode paired dataset, and an open human evaluation platform to find out.

Overview
Overview video
Drop an mp4 at assets/media/demo_video.mp4 to enable.
Abstract

Humans imitate at the level of intent: given a demonstration, we infer its goal and carry it out with whatever tools, objects, and layouts are at hand. Current robot policies instead learn observation-to-action mappings from visual inputs and language instructions, without explicitly inferring the demonstrated task. Learning from human video thus remains largely trajectory-level: models can replay motions in near-identical scenes, but still struggle to imitate what the demonstrator intends rather than merely what they do. We introduce The Imitator Game, a four-level benchmark (L0–L3) that progressively widens the gap between the human demonstration and the robot's own scene, isolating where trajectory replay ceases to suffice and task understanding becomes necessary. We pair it with IG-10K, the largest environment-aligned paired human–robot dataset to date and the only one instantiated across all four levels in both real and simulated settings (20,000+ paired episodes, 50+ tasks, 6 domains), and Imitator Arena, an open platform for blind A/B human evaluation. Across nine state-of-the-art models, performance is stable from L0 to L2 but collapses at L3, identifying functional substitution — achieving the same intent through a different object affordance — as the decisive barrier to intent-level imitation.

Imitator Game teaser: demonstrator scene paired with imitator scene and Arena evaluation panel
The Imitator Game. (Left) A human demonstrates a manipulation task in the Demonstrator Scene (green), and a robot watches the video and reproduces the underlying intent in the Imitator Scene (orange). (Right) The Imitator Arena: human evaluators make blind A/B comparisons of two anonymised model rollouts against the demonstrator video.
Four levels

Increasing abstraction of imitation

Each task is instantiated at four levels of progressive mismatch between the human demonstration and the robot's scene — from exact replay (L0) to intent-level affordance substitution (L3).

REF Human Reference

Every comparison starts from a human demonstrator video recorded in the Demonstrator Scene. This reference video is shown to the robot imitator — and to human Arena evaluators — as the sole specification of what should be reproduced. There is no separate task description, goal predicate, or symbolic instruction.

Example A human steeps a tea bag in hot water using the teapot and cup on the table. The robot must watch this clip and reproduce the intent.
3× Speed
Human reference video assets/media/levels/human_ref.mp4

L0 Trajectory Imitation

The imitator scene is identical to the demonstrator scene — same objects, same layout. The policy is expected to reproduce the demonstrated trajectory.

Example Same tea bag, teapot and cup in the same positions; the imitator pours along essentially the demonstrated motion.
6× Speed
L0 demo video assets/media/levels/l0.mp4

L1 Object End-State Imitation

Same objects, but their spatial configuration has changed. A literal replay of the demonstrated trajectory would fail. The policy must achieve the same final object states via a different path.

Example Same teapot and cup, but the tea bag has moved to the front; the imitator should adjust its movements to steep the tea bag into the cup.
6× Speed
L1 demo video assets/media/levels/l1.mp4

L2 Semantic Task Imitation

At least one task-relevant object has changed in appearance or geometry, but its semantic category is preserved. The precise final states are no longer reproducible, but the semantic task remains achievable.

Example Demonstrator uses a tea bag to steep tea; imitator's scene has a bowl of tea leaves instead. The imitator must infer the underlying task of "brew tea".
6× Speed
L2 demo video assets/media/levels/l2.mp4

L3 Affordance-Adapted Imitation

The demonstrated object is replaced by one of different semantics and function. The policy must infer the underlying intent and re-purpose a different affordance to achieve the same goal.

Example No tea is present — the imitator must serve the same intent by mixing matcha powder instead of steeping a tea bag.
6× Speed
L3 demo video assets/media/levels/l3.mp4
Four imitation levels
Four levels of imitation on a single tea-making demonstration. L0 same objects and layout (trajectory replay); L1 rearranged layout (adapt trajectory); L2 tea leaves instead of tea bag (semantic task); L3 matcha powder instead of tea (affordance substitution).
IG-10K dataset

Paired human–robot dataset at scale

The largest environment-aligned human–robot paired manipulation dataset to date, spanning real-world episodes, interactive simulation environments, and the full task distribution across six domains — with dense offline annotations.

20K+ Paired episodes
11.7K+ Real paired clips
200 Simulation envs
4 Imitation levels
50+ Base tasks
340+ Objects
IG-10K dataset overview
The IG-10K dataset. (a) 20,000+ paired episodes spanning household, supermarket, restaurant, logistics, hospital, and laboratory tasks. (b) 340+ real and simulated objects instantiating the L0–L3 scene variations. (c) Dense annotations per episode: semantic masks, 3D MANO hand poses, depth images, and multi-abstraction language descriptions.
Experiment

Imitator trajectories: Training & Test tasks

For each evaluation task, a human reference video is followed by representative imitator trajectories under increasing scene mismatch (L0 → L3). The five training tasks were included during IG-10K pretraining; the five test tasks are held-out and evaluate zero-shot transfer, from-scratch few-shot learning, and pretrain+finetune (P+FT) adaptation. Seen-task evaluation perturbs object placements across hierarchy levels, so success requires robust imitation rather than fixed-layout replay.

Experiment tasks overview
The experiment. Evaluation spans training tasks (left) and held-out test tasks (right). For each task, the first image shows the human reference video, followed by representative imitator trajectories under increasing scene mismatch from L0 to L3. The real-world evaluation covers seen-task imitation, unseen zero-shot transfer, few-shot adaptation, and Arena human judgments.
Results

Experimental findings

We evaluate 9 state-of-the-art models (15 trained variants) across simulation, real-world deployment, and human Arena judgments — along three axes.

Q1

Video beats text conditioning

Video-based methods lead on seen tasks. ACT/DINOv2 reaches SR = 0.81 in simulation; XSkill wins the real-world Arena with SR = 0.63, WR = 0.89.

Q2

Data scaling is promising

Models falls below 13% zero-shot SR on unseen tasks in simulation. IG-10K pretraining unlocks few-shot gains — with benefits growing with pretraining scale.

Q3

L3 remains a decisive barrier

Performance is stable from L0 to L2 (~0.40 avg SR), then drops to 0.29 at L3. Functional substitution — achieving intent via different affordance — is unsolved frontier.

Try it

Judge the imitators yourself

The Arena shows a human reference video and two anonymous model rollouts on the same scene. Pick which one better imitates the demonstrated intent. Your votes are recorded anonymously to help rank model performance.

Citation

BibTeX

@unpublished{imitatorgame2026,
  title     = {The Imitator Game: Benchmarking Robot Imitative Ability Beyond Action Prediction},
  author    = {Anonymous Author(s)},
  note      = {Submitted to CoRL 2026, under review},
  year      = {2026}
}