Live cart-pole arena
Use ← and → in manual mode.
Live training dashboard
Statistics update after every episode.0
Episode
0.00 s
Longest balance
0.00
Average reward
0%
Success rate
1.000
Exploration rate
0
Training steps
0.0000
Learning loss
—
Latest failure
| # | Algorithm | Duration | Reward | Steps | Result | Failure |
|---|
0.25×1×4×
0 / 0
Replay frame
VS
Choose two strategies to compare average duration, reward, and success rate.
Mode and agent
Current decision
Selected actionNone
Confidence0%
Decision typeWaiting
Expected return0.00
Important signalPole angle
ReasonStart an episode.
Decision trace ready.
Experiment configuration
Changes apply when the next episode starts.
Learning controls
0.0010.0800.5
0.50.9700.999
01.001
0.90.99201
00.030.3
Training schedule
Physics settings
39.820 m/s²
0.250.75 m1.5
0.030.15 kg1
0.31.0 kg3
210.0 N30
24.8 m8
Environment variation
0.5°±5.0°18°
00.0000.2
00 steps12
00.0150.25
Reward design
Challenge target
5 s30 s120 s
8°24°45°
Current target progress0%
Progressive challenges
Achievements
0 unlockedData, model, and reporting
Progress is stored locally in your browser. Exports contain no server-side data.
Learning guide
State, action, and reward
The state describes cart position, velocity, pole angle, and angular velocity. The agent chooses a force, then receives a reward.
Exploration and exploitation
Exploration tests unfamiliar actions. Exploitation chooses actions currently expected to produce the highest future reward.
Discount factor
A higher discount factor values future balance more strongly. A lower value emphasizes immediate rewards.
Browser algorithms
Tabular methods use discretized states. DQN-style modes use lightweight linear value approximators suitable for an interactive browser lesson.