robomotic/lewm-breakout-plays

LeWM Breakout Plays Dataset

This repository contains Atari Breakout trajectories generated by various heuristic policies for the LeWorldModel (LeWM) project. The datasets are organized by the heuristic policy that generated them, and each contains the corresponding Parquet files and statistics.

Available Heuristics

• center: A policy that constantly tries to keep the paddle in the exact center of the screen. This generates data with minimal paddle variance, useful for causal inference on object presence.
• heuristic: An expert-level algorithmic policy that tracks the ball's X position dynamically to successfully hit it. This yields high-reward, long-episode sequences.
• passive: A completely inactive policy that performs no moving actions, generating sequences that show how the game evolves purely passively (without player interaction).
• random_active: A policy that selects actions randomly but ensures the game starts (by occasionally pressing FIRE). Generates diverse, short-lived chaotic trajectories.
• wall_hugger: A policy that drives the paddle to one of the walls and stays there. This creates a data distribution where the paddle's spatial location is extremely biased.

---

🔬 Representation Quality Evaluation

These datasets are used to train and evaluate the LeWorldModel (LeWM), a JEPA-based world model for Atari Breakout. The quality of the learned representations is measured via MLP probing on frozen latent embeddings from the trained encoder and its causal predictor.

Probe Procedure

After training, lightweight MLP probes (3-layer: 256→128→output) and linear probes are trained on top of frozen encoder/predictor representations. No gradient flows back through the JEPA model. Each probe addresses a specific question about what physics the encoder has absorbed:

Task	Input	Target	Why
ball_pos	Encoder CLS at $t$	Ball $(x,y)$ at $t$	Does the encoder localise the ball?
ball_vel	Encoder CLS at $t$	Ball $(v_x,v_y)$ at $t$	Does the encoder track motion?
ball_pos_pred	ARPredictor output at $t$	Ball $(x,y)$ at $t+1$	Does the world model predict next-frame position?
ball_vel_pred	ARPredictor output at $t$	Ball $(v_x,v_y)$ at $t+1$	Does the predictor encode implicit momentum?
paddle_pos	Encoder CLS at $t$	Paddle $x$ at $t$	Does the encoder locate the controlled paddle?
paddle_pos_act	CLS + action at $t$	Paddle $x$ at $t$	How much does action conditioning improve paddle prediction?

Baselines include: (1) a static persistence predictor that copies the current label as the prediction; (2) a randomly-initialised ViT encoder (sanity check — should score near R²≈0); (3) a patch-mean alternative to the CLS token; and (4) linear probes to measure linear separability.

Probe Results — 100-Epoch Reference Run

Evaluation split: held-out my_datasets_eval/ (≈13,225 sequences from heuristic + random_active policies, never seen during training).

MLP Probe — R² (test set)

Task	ep 1	ep 10	ep 25	ep 50	ep 100
ball_pos	+0.178	+0.306	+0.212	+0.234	+0.272
ball_vel	-0.222	-0.314	-0.269	-0.296	-0.271
ball_pos_pred	+0.131	+0.243	+0.230	+0.226	+0.253
ball_vel_pred	-0.144	-0.299	-0.260	-0.249	-0.270
paddle_pos	-0.253	-0.548	-0.370	-0.562	+0.319
paddle_pos_act	-0.169	-0.170	-0.187	-0.536	+0.456

Linear Probe — R² (test set)

Task	ep 1	ep 10	ep 25	ep 50	ep 100
ball_pos	-0.143	+0.244	+0.249	+0.327	+0.301
ball_vel	-0.022	-0.062	-0.059	-0.094	-0.078
ball_pos_pred	-0.194	+0.198	+0.282	+0.278	+0.276
ball_vel_pred	-0.017	-0.031	-0.040	-0.058	-0.068
paddle_pos	-0.384	-0.173	-0.110	-0.009	+0.478
paddle_pos_act	-0.076	+0.077	+0.143	+0.183	+0.607

Sanity-Check Baselines

Baseline	ball_pos R²	ball_vel R²	paddle_pos R²
Static copy (predict $t$ as $t+1$)	+0.955	+0.821	+0.778
Random encoder	-0.351	-0.018	-0.034

Key Findings

• Ball position is learned early: R² reaches +0.31 by epoch 10. The linear probe improves monotonically, confirming representations become more linearly decodable over the full 100 epochs.
• Ball velocity is not encoded: Negative R² across all epochs and probe types — JEPA training does not provide an explicit signal to retain first-order dynamics.
• Predictor tracks ball position well: At epoch 100, the ARPredictor R² (+0.25) is within 2% of the encoder R² (+0.27), demonstrating genuine next-frame anticipation.
• Paddle position emerges late: Val R² grows steadily (reaching +0.92 at ep100), but test R² is negative until epoch 100 when it jumps to +0.32/+0.48 — a distribution-shift effect from the held-out policy mix.
• Action conditioning is effective: Adding the action embedding boosts paddle prediction by +0.13 (linear, ep100), confirming the ARPredictor correctly absorbs action-conditioned dynamics.
• Random encoder is uninformative: R² ≈ −0.35 for ball_pos under a random ViT confirms all probe gains come from training, not pixel statistics.

For full methodology, all results, and reproduction instructions, see the companion model repository at robomotic/lewm-breakout.