docs/BENCHMARKS.md

SLM-Lab Benchmarks

Reproducible deep RL algorithm validation across Gymnasium environments (Classic Control, Box2D, MuJoCo, Atari).

---

Usage

After installation, copy SPEC_FILE and SPEC_NAME from result tables below (Atari uses one shared spec file - see Phase 4).

Running Benchmarks

Local - runs on your machine (Classic Control: minutes):

slm-lab run SPEC_FILE SPEC_NAME train

Remote - cloud GPU via dstack, auto-syncs to HuggingFace:

source .env && slm-lab run-remote --gpu SPEC_FILE SPEC_NAME train -n NAME

Remote setup: cp .env.example .env then set HF_TOKEN. See README for dstack config.

Atari

All games share one spec file (54 tested, 5 hard exploration skipped). Use -s env=ENV to substitute. Runs take ~2-3 hours on GPU.

source .env && slm-lab run-remote --gpu -s env=ALE/Pong-v5 slm_lab/spec/benchmark/ppo/ppo_atari.json ppo_atari train -n pong

Download Results

Trained models and metrics sync to HuggingFace. Pull locally:

source .env && slm-lab pull SPEC_NAME
slm-lab list  # see available experiments

Benchmark Contribution

To ensure benchmark integrity, follow these steps when adding or updating results:

1. Audit Spec Settings

• Before Running: Ensure spec.json matches the Settings line defined in each benchmark table.
• Example: max_frame 3e5 | num_envs 4 | max_session 4 | log_frequency 500
• After Pulling: Verify the downloaded spec.json matches these rules before using the data.

2. Run Benchmark & Commit Specs

• Run: Execute the benchmark locally or remotely using the commands in Usage.
• Commit Specs: Always commit the spec.json file used for the run to the repo.
• Table Entry: Ensure BENCHMARKS.md has an entry with the correct SPEC_FILE and SPEC_NAME.

3. Record Scores & Plots

• Score: At run completion, extract total_reward_ma from logs (trial_metrics).
• Link: Add HuggingFace folder link: [FOLDER](https://huggingface.co/datasets/SLM-Lab/benchmark-dev/tree/main/data/FOLDER)
• Pull data: source .env && uv run hf download SLM-Lab/benchmark-dev --include "data/FOLDER/*" --local-dir hf_data --repo-type dataset
• Plot: Generate with folders from table:

  slm-lab plot -t "CartPole-v1" -f ppo_cartpole_2026...,dqn_cartpole_2026...
  

Environment Settings

Standardized settings for fair comparison. The Settings line in each result table shows these values.

Env Category	num_envs	max_frame	log_frequency	grace_period
Classic Control	4	2e5-3e5	500	1e4
Box2D	8	3e5	1000	5e4
MuJoCo	16	1e6-10e6	1e4	1e5-1e6
Atari	16	10e6	10000	5e5

Hyperparameter Search

When algorithm fails to reach target, run search instead of train:

slm-lab run SPEC_FILE SPEC_NAME search                                        # local
source .env && slm-lab run-remote --gpu SPEC_FILE SPEC_NAME search -n NAME    # remote

Stage	Mode	Config	Purpose
ASHA	search	max_session=1, search_scheduler enabled	Wide exploration with early stopping
Multi	search	max_session=4, NO search_scheduler	Robust validation with averaging
Validate	train	Final spec	Confirmation run

Do not use search result in benchmark results - use final validation run with committed spec.

Search budget: ~3-4 trials per dimension (8 trials = 2-3 dims, 16 = 3-4 dims, 20+ = 5+ dims).

{
  "meta": {
    "max_session": 1, "max_trial": 16,
    "search_resources": {"cpu": 1, "gpu": 0.125},
    "search_scheduler": {"grace_period": 1e5, "reduction_factor": 3}
  },
  "search": {
    "agent.algorithm.gamma__uniform": [0.98, 0.999],
    "agent.algorithm.lam__uniform": [0.9, 0.98],
    "agent.net.optim_spec.lr__loguniform": [1e-4, 1e-3]
  }
}

Progress

Phase	Category	Envs	REINFORCE	SARSA	DQN	DDQN+PER	A2C	PPO	SAC	Overall
1	Classic Control	3	🔄	🔄	🔄	🔄	🔄	🔄	🔄	Rerun pending
2	Box2D	2	N/A	N/A	🔄	🔄	🔄	🔄	🔄	Rerun pending
3	MuJoCo	11	N/A	N/A	N/A	N/A	🔄	🔄	🔄	Rerun pending
4	Atari	59	N/A	N/A	Skip	Skip	Skip	🔄	N/A	54 games (not in this rerun)

Legend: ✅ Solved | ⚠️ Close (>80%) | 📊 Acceptable | ❌ Failed | 🔄 In progress/Pending | Skip Not started | N/A Not applicable

---

Results

Phase 1: Classic Control

1.1 CartPole-v1

Docs: CartPole | State: Box(4) | Action: Discrete(2) | Target reward MA > 400

Settings: max_frame 2e5 | num_envs 4 | max_session 4 | log_frequency 500

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
REINFORCE	✅	469.68	slm_lab/spec/benchmark/reinforce/reinforce_cartpole.json	reinforce_cartpole	reinforce_cartpole_2026_01_30_215510
SARSA	✅	421.58	slm_lab/spec/benchmark/sarsa/sarsa_cartpole.json	sarsa_boltzmann_cartpole	sarsa_boltzmann_cartpole_2026_01_30_215508
DQN	⚠️	188.07	slm_lab/spec/benchmark/dqn/dqn_cartpole.json	dqn_boltzmann_cartpole	dqn_boltzmann_cartpole_2026_01_30_215213
DDQN+PER	✅	432.88	slm_lab/spec/benchmark/dqn/dqn_cartpole.json	ddqn_per_boltzmann_cartpole	ddqn_per_boltzmann_cartpole_2026_01_30_215454
A2C	✅	499.73	slm_lab/spec/benchmark/a2c/a2c_gae_cartpole.json	a2c_gae_cartpole	a2c_gae_cartpole_2026_01_30_215337
PPO	✅	499.54	slm_lab/spec/benchmark/ppo/ppo_cartpole.json	ppo_cartpole	ppo_cartpole_2026_01_30_221924
SAC	⚠️	359.69	slm_lab/spec/benchmark/sac/sac_cartpole.json	sac_cartpole	sac_cartpole_2026_01_30_221934

1.2 Acrobot-v1

Docs: Acrobot | State: Box(6) | Action: Discrete(3) | Target reward MA > -100

Settings: max_frame 3e5 | num_envs 4 | max_session 4 | log_frequency 500

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
DQN	✅	-94.81	slm_lab/spec/benchmark/dqn/dqn_acrobot.json	dqn_boltzmann_acrobot	dqn_boltzmann_acrobot_2026_01_30_215429
DDQN+PER	✅	-85.17	slm_lab/spec/benchmark/dqn/ddqn_per_acrobot.json	ddqn_per_acrobot	ddqn_per_acrobot_2026_01_30_215436
A2C	✅	-83.75	slm_lab/spec/benchmark/a2c/a2c_gae_acrobot.json	a2c_gae_acrobot	a2c_gae_acrobot_2026_01_30_215413
PPO	✅	-81.43	slm_lab/spec/benchmark/ppo/ppo_acrobot.json	ppo_acrobot	ppo_acrobot_2026_01_30_215352
SAC	✅	-97.08	slm_lab/spec/benchmark/sac/sac_acrobot.json	sac_acrobot	sac_acrobot_2026_01_30_215401

1.3 Pendulum-v1

Docs: Pendulum | State: Box(3) | Action: Box(1) | Target reward MA > -200

Settings: max_frame 3e5 | num_envs 4 | max_session 4 | log_frequency 500

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
A2C	❌	-553.00	slm_lab/spec/benchmark/a2c/a2c_gae_pendulum.json	a2c_gae_pendulum	a2c_gae_pendulum_2026_01_30_215421
PPO	✅	-168.26	slm_lab/spec/benchmark/ppo/ppo_pendulum.json	ppo_pendulum	ppo_pendulum_2026_01_30_215944
SAC	✅	-152.33	slm_lab/spec/benchmark/sac/sac_pendulum.json	sac_pendulum	sac_pendulum_2026_01_30_215454

Phase 2: Box2D

2.1 LunarLander-v3 (Discrete)

Docs: LunarLander | State: Box(8) | Action: Discrete(4) | Target reward MA > 200

Settings: max_frame 3e5 | num_envs 8 | max_session 4 | log_frequency 1000

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
DQN	⚠️	183.64	slm_lab/spec/benchmark/dqn/dqn_lunar.json	dqn_concat_lunar	dqn_concat_lunar_2026_01_30_215529
DDQN+PER	✅	261.49	slm_lab/spec/benchmark/dqn/ddqn_per_lunar.json	ddqn_per_concat_lunar	ddqn_per_concat_lunar_2026_01_30_215532
A2C	❌	9.53	slm_lab/spec/benchmark/a2c/a2c_gae_lunar.json	a2c_gae_lunar	a2c_gae_lunar_2026_01_30_215529
PPO	⚠️	159.02	slm_lab/spec/benchmark/ppo/ppo_lunar.json	ppo_lunar	ppo_lunar_2026_01_30_215550
SAC	❌	-75.43	slm_lab/spec/benchmark/sac/sac_lunar.json	sac_lunar	sac_lunar_2026_01_30_215552

2.2 LunarLander-v3 (Continuous)

Docs: LunarLander | State: Box(8) | Action: Box(2) | Target reward MA > 200

Settings: max_frame 3e5 | num_envs 8 | max_session 4 | log_frequency 1000

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
A2C	❌	-38.18	slm_lab/spec/benchmark/a2c/a2c_gae_lunar.json	a2c_gae_lunar_continuous	a2c_gae_lunar_continuous_2026_01_30_215630
PPO	⚠️	165.48	slm_lab/spec/benchmark/ppo/ppo_lunar.json	ppo_lunar_continuous	ppo_lunar_continuous_2026_01_31_104549
SAC	✅	208.60	slm_lab/spec/benchmark/sac/sac_lunar.json	sac_lunar_continuous	sac_lunar_continuous_2026_01_31_104537

Phase 3: MuJoCo

Docs: MuJoCo environments | State/Action: Continuous | Target: Practical baselines (no official "solved" threshold)

Settings: max_frame 4e6-10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm: PPO only - SAC omitted (off-policy = heavy compute for systematic benchmarking). Network: MLP [256,256] tanh, orthogonal init.

Spec Variants: Two unified specs in ppo_mujoco.json, plus individual specs for edge cases.

SPEC_NAME	Envs	Key Config
ppo_mujoco	HalfCheetah, Walker, Humanoid, HumanoidStandup	gamma=0.99, lam=0.95
ppo_mujoco_longhorizon	Reacher, Pusher	gamma=0.997, lam=0.97
Individual specs	Hopper, Swimmer, Ant, IP, IDP	See spec files for tuned hyperparams

Reproduce: Copy ENV, SPEC_FILE, SPEC_NAME from table. Use -s max_frame= for all specs, add -s env= for unified specs:

# Unified specs (ppo_mujoco.json)
source .env && slm-lab run-remote --gpu -s env=ENV -s max_frame=MAX_FRAME \
  slm_lab/spec/benchmark/ppo/ppo_mujoco.json SPEC_NAME train -n NAME

# Individual specs (env hardcoded)
source .env && slm-lab run-remote --gpu -s max_frame=MAX_FRAME \
  slm_lab/spec/benchmark/ppo/SPEC_FILE SPEC_NAME train -n NAME

ENV	MAX_FRAME	SPEC_FILE	SPEC_NAME
HalfCheetah-v5	10e6	ppo_mujoco.json	ppo_mujoco
Walker2d-v5	10e6	ppo_mujoco.json	ppo_mujoco
Humanoid-v5	10e6	ppo_mujoco.json	ppo_mujoco
HumanoidStandup-v5	4e6	ppo_mujoco.json	ppo_mujoco
Hopper-v5	4e6	ppo_hopper.json	ppo_hopper
Swimmer-v5	4e6	ppo_swimmer.json	ppo_swimmer
Ant-v5	10e6	ppo_ant.json	ppo_ant
Reacher-v5	4e6	ppo_mujoco.json	ppo_mujoco_longhorizon
Pusher-v5	4e6	ppo_mujoco.json	ppo_mujoco_longhorizon
InvertedPendulum-v5	4e6	ppo_inverted_pendulum.json	ppo_inverted_pendulum
InvertedDoublePendulum-v5	10e6	ppo_inverted_double_pendulum.json	ppo_inverted_double_pendulum

3.1 Hopper-v5

Docs: Hopper | State: Box(11) | Action: Box(3) | Target reward MA > 2000

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	⚠️	1174.57	slm_lab/spec/benchmark/ppo/ppo_hopper.json	ppo_hopper	ppo_hopper_2026_01_30_220138

3.2 HalfCheetah-v5

Docs: HalfCheetah | State: Box(17) | Action: Box(6) | Target reward MA > 5000

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	5851.70	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco	ppo_mujoco_halfcheetah_2026_01_30_230302

3.3 Walker2d-v5

Docs: Walker2d | State: Box(17) | Action: Box(6) | Target reward MA > 3500

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	4042.07	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco	ppo_mujoco_walker2d_2026_01_30_222124

3.4 Ant-v5

Docs: Ant | State: Box(105) | Action: Box(8) | Target reward MA > 2000

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	2514.64	slm_lab/spec/benchmark/ppo/ppo_ant.json	ppo_ant	ppo_ant_2026_01_31_042006

3.5 Swimmer-v5

Docs: Swimmer | State: Box(8) | Action: Box(2) | Target reward MA > 200

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	229.31	slm_lab/spec/benchmark/ppo/ppo_swimmer.json	ppo_swimmer	ppo_swimmer_2026_01_30_215922

3.6 Reacher-v5

Docs: Reacher | State: Box(11) | Action: Box(2) | Target reward MA > -10

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	-5.08	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco_longhorizon	ppo_mujoco_longhorizon_reacher_2026_01_30_215805

3.7 Pusher-v5

Docs: Pusher | State: Box(23) | Action: Box(7) | Target reward MA > -50

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	-49.09	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco_longhorizon	ppo_mujoco_longhorizon_pusher_2026_01_30_215824

3.8 InvertedPendulum-v5

Docs: InvertedPendulum | State: Box(4) | Action: Box(1) | Target reward MA ~1000

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	944.87	slm_lab/spec/benchmark/ppo/ppo_inverted_pendulum.json	ppo_inverted_pendulum	ppo_inverted_pendulum_2026_01_30_230211

3.9 InvertedDoublePendulum-v5

Docs: InvertedDoublePendulum | State: Box(11) | Action: Box(1) | Target reward MA ~8000

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	7622.00	slm_lab/spec/benchmark/ppo/ppo_inverted_double_pendulum.json	ppo_inverted_double_pendulum	ppo_inverted_double_pendulum_2026_01_30_220651

3.10 Humanoid-v5

Docs: Humanoid | State: Box(376) | Action: Box(17) | Target reward MA > 1000

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	3774.08	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco	ppo_mujoco_humanoid_2026_01_30_222339

3.11 HumanoidStandup-v5

Docs: HumanoidStandup | State: Box(376) | Action: Box(17) | Target reward MA > 100000

Settings: max_frame 4e6 | num_envs 16 | max_session 4 | log_frequency 1e4

Algorithm	Status	MA	SPEC_FILE	SPEC_NAME	HF Repo
PPO	✅	165841.17	slm_lab/spec/benchmark/ppo/ppo_mujoco.json	ppo_mujoco	ppo_mujoco_humanoidstandup_2026_01_30_215802

Phase 4: Atari

Docs: Atari environments | State: Box(84,84,4 after preprocessing) | Action: Discrete(4-18, game-dependent) | Solved: Game-specific thresholds

Settings: max_frame 10e6 | num_envs 16 | max_session 4 | log_frequency 10000

Environment:

• Gymnasium ALE v5 with life_loss_info=true
• v5 is harder than v4 due to sticky actions (default repeat_action_probability=0.25 vs v4's 0.0), which randomly repeats agent actions to simulate console stochasticity and prevent memorization, following Machado et al. (2018) research best practices. See ALE version history.

Algorithm: PPO:

• Network: ConvNet [32,64,64] + 512fc (Nature CNN), orthogonal init, normalize=true, clip_grad_val=0.5
• Hyperparams: AdamW (lr=2.5e-4, eps=1e-5), minibatch_size=256, time_horizon=128, training_epoch=4, clip_eps=0.1, entropy_coef=0.01

Lambda Variants: All use one spec file (slm_lab/spec/benchmark/ppo/ppo_atari.json), differing only in GAE lambda. Lower lambda = bias toward immediate rewards (action games), higher = longer credit horizon (strategic games).

SPEC_NAME	Lambda	Best for
ppo_atari	0.95	Long-horizon, strategic games (default)
ppo_atari_lam85	0.85	Mixed/moderate games
ppo_atari_lam70	0.70	Fast action games

Reproduce: Copy ENV from first column, SPEC_NAME from column header. All use the same SPEC_FILE:

source .env && slm-lab run-remote --gpu -s env=ENV \
  slm_lab/spec/benchmark/ppo/ppo_atari.json SPEC_NAME train -n NAME

ENV\SPEC_NAME	ppo_atari	ppo_atari_lam85	ppo_atari_lam70
ALE/Adventure-v5	Skip	Skip	Skip
ALE/AirRaid-v5	8245	-	-
ALE/Alien-v5	1453	1353	1274
ALE/Amidar-v5	574	580	-
ALE/Assault-v5	4059	4293	3314
ALE/Asterix-v5	2967	3482	-
ALE/Asteroids-v5	1497	1554	-
ALE/Atlantis-v5	792886	754k	710k
ALE/BankHeist-v5	1045	1045	-
ALE/BattleZone-v5	21270	26383	13857
ALE/BeamRider-v5	2765	-	-
ALE/Berzerk-v5	1072	-	-
ALE/Bowling-v5	46.45	-	-
ALE/Boxing-v5	91.17	-	-
ALE/Breakout-v5	191	292	327
ALE/Carnival-v5	3071	3013	3967
ALE/Centipede-v5	3917	-	4915
ALE/ChopperCommand-v5	5355	-	-
ALE/CrazyClimber-v5	107183	107370	-
ALE/Defender-v5	37162	-	51439
ALE/DemonAttack-v5	7755	-	16558
ALE/DoubleDunk-v5	-2.38	-	-
ALE/ElevatorAction-v5	5446	363	3933
ALE/Enduro-v5	414	898	872
ALE/FishingDerby-v5	22.80	27.10	-
ALE/Freeway-v5	31.30	-	-
ALE/Frostbite-v5	301	275	267
ALE/Gopher-v5	4172	-	6508
ALE/Gravitar-v5	599	253	145
ALE/Hero-v5	21052	28238	-
ALE/IceHockey-v5	-3.93	-5.58	-7.36
ALE/Jamesbond-v5	662	-	-
ALE/JourneyEscape-v5	-1582	-1252	-1547
ALE/Kangaroo-v5	2623	9912	-
ALE/Krull-v5	7841	-	-
ALE/KungFuMaster-v5	18973	28334	29068
ALE/MontezumaRevenge-v5	Skip	Skip	Skip
ALE/MsPacman-v5	2308	2372	2297
ALE/NameThisGame-v5	5993	-	-
ALE/Phoenix-v5	7940	-	15659
ALE/Pitfall-v5	Skip	Skip	Skip
ALE/Pong-v5	15.01	16.91	12.85
ALE/Pooyan-v5	4704	-	5716
ALE/PrivateEye-v5	Skip	Skip	Skip
ALE/Qbert-v5	15094	-	-
ALE/Riverraid-v5	7319	9428	-
ALE/RoadRunner-v5	24204	37015	-
ALE/Robotank-v5	20.07	8.24	2.59
ALE/Seaquest-v5	1796	-	-
ALE/Skiing-v5	-19340	-22980	-29975
ALE/Solaris-v5	2094	-	-
ALE/SpaceInvaders-v5	726	-	-
ALE/StarGunner-v5	31862	-	47495
ALE/Surround-v5	-2.52	-	-6.79
ALE/Tennis-v5	-7.66	-4.41	-
ALE/TimePilot-v5	4668	-	-
ALE/Tutankham-v5	203	217	-
ALE/UpNDown-v5	182472	-	-
ALE/Venture-v5	Skip	Skip	Skip
ALE/VideoPinball-v5	31385	-	56746
ALE/WizardOfWor-v5	5814	5466	4740
ALE/YarsRevenge-v5	17120	-	-
ALE/Zaxxon-v5	10756	-	-

Legend: Bold = Best score | Skip = Hard exploration | - = Not tested

---

Sticky Actions Validation (v5 vs v4-style)

Testing hypothesis that lower scores are due to sticky actions (repeat_action_probability=0.25 in v5 vs 0.0 in v4/CleanRL).

Environment: Same as above, but with repeat_action_probability=0.0 (matching CleanRL/old v4 behavior)

Reproduce: Copy ENV from first column:

source .env && slm-lab run-remote --gpu -s env=ENV \
  slm_lab/spec/benchmark/ppo/ppo_atari.json ppo_atari_nosticky train -n NAME

Results (Testing games with significant regression):

ENV	v5 (sticky=0.25)	v4-style (sticky=0.0)	Diff	% Change
ALE/Skiing-v5	-19340	-	-	-
ALE/Frostbite-v5	301	-	-	-
ALE/ElevatorAction-v5	5446	-	-	-
ALE/Gravitar-v5	599	-	-	-
ALE/WizardOfWor-v5	5814	-	-	-
ALE/Alien-v5	1453	-	-	-
ALE/KungFuMaster-v5	29068	-	-	-
ALE/Atlantis-v5	792886	-	-	-
ALE/Pong-v5	15.01	-	-	-
ALE/Breakout-v5	191	-	-	-