Games

Game rules, rewards and encoders are separate components. The table describes the built-in defaults; constructors expose the configurable details.

Game	Players	Decisions	Chance	Information	Actions	Observation	Reward arguments (defaults)	Adapter
Backgammon	2	Sequential	Dice	Perfect	1,352 discrete ids	Fixed CHW tensor	win=1, gammon=2, backgammon=3	PettingZoo AEC
Chess	2	Sequential	No	Perfect	AlphaZero 8x8x73	Selectable fixed CHW encoder	win=1, loss=-1, draw=0	PettingZoo AEC
Connect 4	2	Sequential	No	Perfect	7 columns	Fixed CHW tensor	win=1, loss=-1, draw=0	PettingZoo AEC
GridWorld	1	Sequential	No	Perfect	4 directions	Fixed CHW tensor	step=0, goal=1	Gymnasium
CarRacing	1	Sequential	Track-generation seed (root only)	Perfect	5 discrete controls	Rendered CHW pixels (3, 96, 96)	tile=1000, step=-0.1, off_playfield=-100	None (withheld until the HWC adapter presentation lands)
Kuhn poker	2-10	Sequential	Card deal	Imperfect	Pass / bet	Fixed information-state tensor	scale=1	PettingZoo AEC
Leduc poker	2	Sequential	Card deal	Imperfect	Fold / call / raise	Fixed information-state tensor	scale=1	PettingZoo AEC
Snake	2-8	Simultaneous	Placement and respawn	Perfect	3 relative moves	Egocentric fixed CHW tensor	step=0, food=0, loss=-1, draw=0, kill=0, win=1, survival=0	PettingZoo Parallel
Texas Hold'em	2-9	Sequential	Deal and board	Imperfect	Fold / call / raise	Egocentric fixed CHW tensor	scale=1	PettingZoo AEC

The action interface is one fixed discrete vocabulary per game; continuous and parameterized actions are outside the current action-space boundary. The Decisions column describes one game-wide mode: a game cannot currently switch between sequential and simultaneous decision phases.

Reward arguments are the keyword weights accepted by rf.Reward for that game; unknown names are rejected at Engine construction. Poker's scale multiplies the native chip-delta event.

Observation encoders

Every game constructor accepts an encoder= handle and otherwise selects its registered default. The encoder owns the network-facing observation and action frame, while the game owns rules and state transitions. Chess currently offers multiple views; the other built-in games each expose one. Size a network from game.observation_space().shape after construction.

game = rf.games.Chess(encoder=rf.encoders.RelativeChess())
obs_shape = game.observation_space().shape  # (19, 8, 8)

Encoder	Game	Observation shape	Constructor	Purpose
Snake	Snake	(5, grid_size, grid_size); (5, 20, 20) by default	rf.encoders.Snake()	Mover-relative own-head, own-body, opponent-head, opponent-body, and food planes.
Connect4	Connect 4	(2, 6, 7)	rf.encoders.Connect4()	Acting-player and opponent piece planes.
MinimalChess	Chess	(19, 8, 8)	rf.encoders.MinimalChess()	Absolute piece, turn, castling, en-passant and clock planes; the Chess default.
RelativeChess	Chess	(19, 8, 8)	rf.encoders.RelativeChess()	Mover-relative board and matching action-head mapping.
OpenSpielChess	Chess	(20, 8, 8)	rf.encoders.OpenSpielChess()	OpenSpiel-compatible observation for parity and benchmarking.
AlphaZeroChess	Chess	(119, 8, 8) by default	rf.encoders.AlphaZeroChess(history_length=8)	Uses 14 * history_length + 7 channels for history, repetition, side, move-count, castling and clock features.
Backgammon	Backgammon	(200, 1, 1)	rf.encoders.Backgammon()	Tesauro-style point, bar, borne-off, turn, and dice features.
TexasHoldem	Texas Hold'em	(2 * num_players + 19, 4, 13); (31, 4, 13) by default	rf.encoders.TexasHoldem()	Player-relative cards, betting state, stacks, and positions.
KuhnPoker	Kuhn poker	(3 * players, 1, 1); (6, 1, 1) by default	rf.encoders.KuhnPoker()	Private card and public betting-history information state.
LeducPoker	Leduc poker	(21, 1, 1)	rf.encoders.LeducPoker()	Private/public cards and two-round betting information state.
GridWorld	GridWorld	(2, size, size); (2, 5, 5) by default	rf.encoders.GridWorld()	Agent-position and goal planes.
CarRacingPixels	CarRacing	(3, 96, 96)	rf.encoders.CarRacingPixels()	Rendered top-down frame, raw 0-255 values, channel-major (gym renders HWC).
CarRacingVec	CarRacing	(1, 1, 21)	rf.encoders.CarRacingVec()	Diagnostic pose/velocity/wheel/progress vector; renderer-free training and physics-vs-visual failure isolation.

Before moving action ids between a network and Env, read the action-frame contract.

CarRacing notes

CarRacing ports Gymnasium's CarRacing-v3 (discrete). Faithful means the action contract, observation content, track-generation algorithm, reward and episode rules, and qualitative driving behavior; trajectories, checkpoints, and scores from the Gymnasium build do not transfer (different physics engine and float paths).

• The rendered HUD score is the canonical Gym score (fixed tile/step constants, off-playfield never shown) regardless of the rf.Reward weights you train with.
• Tile progress counts per-wheel contact-entry events, matching Gym's behavior including its sticky reset-contact lap quirk.
• Observations are channel-major (3, 96, 96) raw 0-255 floats per the framework contract; Gymnasium renders the same content as HWC uint8, so pipelines moving between the two must permute and cast.
• Snapshot bytes carry only validated dynamic values (poses, velocities, controls, joint warm-start impulses, progress); the physics world is reconstructed canonically at restore, which is what keeps hostile snapshot bytes safe. Restore is exact over those serialized values and always resumes the same trajectory, but that trajectory departs from a never-snapshotted run at float rounding order (solver bookkeeping is rebuilt, not restored), so restored collection is deterministic rather than record-identical to an uninterrupted run.
• Search-based algorithms (TreeStrap, AlphaZero) are mechanically supported: after the track is realized the game is a deterministic single-agent planning problem. Under the pixels encoder every search node renders a full frame, so search-based training is far more economical with rf.encoders.CarRacingVec().
• The standard-API adapters withhold car_racing until frames can be presented in Gym's native HWC uint8 layout; use rf.Env directly.

See built-in compatibility for supported workflows and the Python API for constructor parameters.