This supplementary website accompanies the paper “Augmentations for Robust and Efficient Imitation Learning in Streamed Video Games,” published at the Conference on Games 2026. The paper studies whether spatiotemporal augmentations that mimic common streaming artifacts like pixelation, blur, scrubs, and ghosting, can improve the sample efficiency and robustness of imitation learning agents trained from limited gameplay demonstrations. On this website, we provide representative examples of the proposed augmentations, qualitative rollout videos of trained agents, and additional visualizations supporting the results reported in the paper. These materials are intended to complement the quantitative evaluations by illustrating both the streaming artifacts modeled by our method and their impact on agent behavior under normal and degraded streaming conditions.
Compare representative rollouts across Clean, Synthetic, and Real Streaming Noise settings for Game 1: Task 1, Game 1: Task 2, and Game 2: Task 3, organized by demonstration budget and method
Get the code
Read the paper
Representative rollouts
Best episodes are the highest-scoring latest completed rollouts for a setting. Median episodes are selected from the middle of the score distribution for each setting
-
Clean conditions
Evaluation through streaming with no induced lag or introduced visual artifacts
-
Synthetic conditions
Evaluation through streaming with no induced lag, using streaming augmentations to introduce visual artifacts
-
Real Streaming Noise conditions
Evaluation through streaming with induced lag, leading to visual artifacts that the streaming augmentations are modeled after
Clean conditions
Representative median and best rollouts are shown for each available method in this setting
-
Game 1, Task 1 | 12 Milestones
No augmentation | Avg. 76.7% (9.20 / 12 milestones) | 5 seeds
Median episode: 66.7% Best episode: 100% All augmentation | Avg: 85.8% (10.30 / 12 milestones) | 5 seeds
Median episode: 100% Best episode: 100% Game 1, Task 2 | 11 Milestones
No augmentation | Avg: 94.4% (10.38 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100% All augmentation | Avg: 97.5% (10.72 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100% Game 2, Task 3 | 15 Milestones
No augmentation | Avg: 55.2% (8.28 / 15) | 5 seeds
Median episode: 60% Best episode: 100% All augmentation | Avg: 65.1% (9.76 / 15 milestones) | 5 seeds
Median episode: 66.7% Best episode: 100% -
Game 1, Task 1 | 12 Milestones
No augmentation | Avg: 43.3% (5.20 / 12 milestones) | 5 seeds
Median episode: 50.0% Best episode: 50.0% All augmentation | Avg: 84.7% (10.16 / 12 milestones) | 5 seeds
Median episode: 91.7% Best episode: 100% Game 1, Task 2 | 11 Milestones
No augmentation | Avg: 88.7% (9.76 / 11 milestones) | 5 seeds
Median episode: 90.9% Best episode: 100% All augmentation | Avg: 98% (10.78 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100% -
Game 1, Task 1 | 12 Milestones
No augmentation | Avg: 44% (5.28 / 12 milestones) | 5 seeds
Median episode: 41.7% Best episode: 66.7% All augmentation | Avg: 79.2% (9.50 / 12 milestones) | 5 seeds
Median episode: 91.7% Best episode: 100% Game 1, Task 2 | 11 Milestones
No augmentation | Avg: 88% (9.68 / 11 milestones) | 5 seeds
Median episode: 90.9% Best episode: 100% All augmentation | Avg: 96.5% (10.62 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100% -
Game 1, Task 1 | 12 Milestones
No augmentation | Avg: 42.2% (5.06 / 12 milestones) | 5 seeds
Median episode: 41.7% Best episode: 75% All augmentation | Avg: 72.2% (8.66 / 12 milestones) | 5 seeds
Median episode: 83.3% Best episode: 100% Game 1, Task 2 | 11 Milestones
No augmentation | Avg: 36.5% (4.02 / 11 milestones) | 5 seeds
Median episode: 27.3% Best episode: 81.8% All augmentation | Avg: 67.5% (7.42 / 11 milestones) | 5 seeds
Median episode: 63.6% Best episode: 100%
Synthetic conditions
Best episodes are the highest-scoring latest completed rollouts for a setting. Median episodes are selected from the middle of the score distribution for each setting
-
Game 1, Task 1 | 12 Milestones
No augmentation | Avg: 72.3% (8.68 / 12 milestones) | 5 seeds
Median episode: 75% Best episode: 100% Standard augmentation | Avg: 76.5% (9.18 / 12 milestones) | 5 seeds
Median episode: 75% Best episode: 100% Streamingaugmentation | Avg: 87% (10.44 / 12 milestones) | 5 seeds
Median episode: 100% Best episode: 100% All augmentation | Avg: 86.7% (10.40 / 12 milestones) | 5 seeds
Median episode: 91.7% Best episode: 100% Game 1, Task 2 | 11 Milestones
Noaugmentation | Avg: 54.9% (6.04 / 11 milestones) | 5 seeds
Median episode: 72.7% Best episode: 100% Standard augmentation | Avg: 53.5% (5.88 / 11 milestones) | 5 seeds
Median episode: 45.5% Best episode: 100% Streamingaugmentation | Avg: 96% (10.56 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100% All augmentation | Avg: 96.2% (10.58 / 11 milestones) | 5 seeds
Median episode: 100% Best episode: 100%
Real Streaming Noise conditions
Best episodes are the highest-scoring latest completed rollouts for a setting. Median episodes are selected from the middle of the score distribution for each setting
-
Game 1, Task 2 | 12 Milestones
No augmentation | Avg: 44.5% (4.90 / 11 milestones) | 1 seed
Median episode: 36.4% Best episode: 81.8% All augmentation | Avg: 90% (9.90 / 11 milestones) | 1 seed
Median episode: 90.9% Best episode: 100%
Augmentation visualization
Game 1, Task 1
Standard Augmentation: Color Jitter
Standard Augmentation: Random Affine
Streaming Augmentation: Fuzziness
Streaming Augmentation: Ghosting
Streaming Augmentation: Pixelation
Streaming Augmentation: Scrubs
All Streaming Augmentations
Combined: All Augmentations
