Simulator rolloutRecorded observation
SIMULATOR
Loading observation… ABC SIMULATOR · CANON WAM CHECKPOINTS & TASKS
What the model imagined.
What the robot did.
Explore generated futures beside the recorded simulation.
Ten shared worlds per checkpoint and task. Every prediction window. Sixty-world results ↓
Throw plastic bottles into the bin
Absolute EEFBlender48 actions / call
TASK SUCCESS— / 10 worlds
MEAN MAX REWARD—
BUDGET / WORLD22,000 actions
SAVED FORECASTS—
Success requires all four bottles in the bin. Partial reward shows progress toward that goal.
01 / CHOOSE A WORLD
Select a thumbnail to exploreThe same scenes, at each checkpoint.
02 / COMPARE OBSERVATION & FORECAST
World 01 SEED —
WAM predictionGenerated future
GENERATED
FORECAST 001 / 459Loading prediction…0 / 22,000 actions0:00 / 12:28 simulated
05,50011,00016,50022,000 actions
INSPECT A FORECASTof 459
Current frame → +6 → +12 → … → +48 actions
Loading recording cadence…
03 / FOLLOW THE OUTCOME
Reward over the rollout.
Step 2,000Step 1,000
Click the chart to seek. Reward is measured every 50 actions. A value of 1 means the task is solved.
04 / SIXTY-WORLD RESULTS
Same seeds for every model · Blender · 22,000-action budgetHow the 14B OpenWAM compares.
Throw plastic bottles into the bin 60 worlds
| Model | Actions / call | Solved | Mean final reward | Median steps to solve |
|---|---|---|---|---|
| OpenWAM canon 14B · step 57,000 (EMA) | 32 | 41 / 60 | 0.900 | 1,550 |
| pi0.5 | 15 | 44 / 60 | 0.917 | 3,050 |
| DreamZero · step 18,000 | 48 | 44 / 60 | 0.917 | 1,925 |
| ABC-DiT | 30 | 33 / 60 | 0.863 | 4,200 |
| OpenWAM-α · step 22,000 | 32 | 6 / 60 | 0.533 | 5,900 |
| DreamZero canon · step 2,000 | 48 | 0 / 60 | 0.029* | – |
Paired by world, the 14B model is not distinguishable from pi0.5 (12 vs 15 worlds solved only by one model, p = 0.70) or DreamZero 18k (8 vs 11, p = 0.65). 18 of its 19 misses stop at 0.75: the last bottle never reaches the bin. *Mean max reward.
Experiments 6–9 tasks 60 worlds each · solved, mean final reward
| Task | OpenWAM canon 14B | OpenWAM-α · step 24,000 | DreamZero · step 18,000 | pi0.5 task-finetune · step 94,000 |
|---|---|---|---|---|
| Turn the mug right side up | 20 / 60 · 0.519 | 17 / 60 · 0.403 | 1 / 60 · 0.033 | 19 / 60 · 0.458 |
| Sweep away the paper scraps | 10 / 60 · 0.381 | 0 / 60 · 0.011 | 1 / 60 · 0.108 | 0 / 60 · 0.036 |
| Load the plates into the dish rack | 3 / 60 · 0.113 | 2 / 60 · 0.042 | 2 / 60 · 0.071 | 1 / 60 · 0.031 |
| Set up the chess pieces | 0 / 56 · 0.004 | 0 / 50 · 0.000 | 1 / 57 · 0.018 | 0 / 58 · 0.000 |
Sweep is the clear gain: the 14B model solves 10 worlds the others miss, with 0 the other way against OpenWAM-α (p = 0.002) and 1 against DreamZero (p = 0.01). Mug flip is level with OpenWAM-α (11 vs 8, p = 0.65). All 10 sweep successes end with the bin upright and the table clear. pi0.5 (15 actions per call) matches the 14B model on mug flip (9 vs 10 worlds solved only by one model, p = 1) but solves no sweep worlds (0 vs 10, p = 0.002). In 2 of the 3 14B plate successes the plate is still in the gripper at the rack edge, and the single pi0.5 plate success is a plate standing against the outside of the rack, so no model reliably solves plates. Chess counts leave out worlds that crashed the physics engine (4, 5, 3 and 2 for pi0.5); OpenWAM-α also has 5 unfinished chess worlds.
Six more tasks 14B OpenWAM only · 10 worlds each
| Task | Solved | Mean final reward | Mean max reward |
|---|---|---|---|
| Pick from the conveyor | 3 / 10 | 0.567 | 0.567 |
| Sort the Lego blocks | 1 / 10 | 0.891 | 0.902 |
| Hand an item to the other side | 1 / 10 | 0.100 | 0.300 |
| Sort the nuts and bolts | 0 / 10 | 0.026 | 0.148 |
| Pour the beads | 0 / 10 | 0.000 | 0.022 |
| Hang the mug on the mug rack | (8 / 10) | (0.850) | (0.917) |
The mug-rack scorer is unreliable: it also counts a mug still held in the gripper next to the rack, or one standing at its base, so that row is not a valid success rate. On Lego, most worlds finish one block short.
About these recordings Protocol, timing & interpretation
Closed-loop simulation
Each call sees the latest simulator cameras and measured state, predicts an action block (48 actions for DreamZero, 32 for OpenWAM 14B), and executes all of it. Generated video is saved for inspection and is never fed back as the next observation.
Independent prediction windows
Every forecast contains the current frame followed by eight generated frames. “Loop this forecast” isolates one window. “Play rollout” follows the full simulator timeline and selects its corresponding forecast.
Matching the training contract
Absolute canonical base-frame end-effector actions, ABC-source quantile normalization, canonical grippers, and inverse kinematics. The canvas puts the top camera above the left and right wrist cameras.
Read results at the right scope
Each checkpoint or task in the viewer shows ten worlds; the sixty-world tables above use the full evaluations. New recordings appear individually as they finish; the ready count shows the scope of the displayed results. All 22,000 actions are budgeted; the last stored simulator image is at action —.