Predict what happens next.
Reason about contact, mass, friction, and motion. A projectile can roll an object, topple a structure, or fracture a glass barrier.
Launch · dynamics · stabilityFOUNDATION MODELS · ROBOTIC INTERACTION
Embodied
Playground: A Long-Horizon Agentic Robotics Benchmark for Physical, Strategic, and Spatial Reasoning
410task environments
100curated tasks
5evaluation suites
16task types
01 / THE PLAYGROUND
The whole agent, in the loop.
Reasoning meets physical consequences.
A foundation-model agent interprets a goal, reads the scene, writes control code, and acts through a robot. Each action changes what it needs to do next.
Embodied Playground evaluates that complete loop under declared observation and tool constraints. Success requires reaching the goal while respecting task rules, physical limits, and execution budgets.
Read the benchmark framing ↗Choose a workable sequence.
Maintain legal intermediate states, order dependent actions, and revise a plan as ammunition, object positions, and available options change.
Tower · assembly · household sequencesConnect geometry to action.
Track poses and cube faces, align interlocking parts, and find grasps and placements that fit the robot’s reachable workspace.
Cube · Lock · bookshelf arrangementA shared interaction interface
Embodiment-specific actions. Independent evaluation.describeRead the task, robot, and action contract.
observeReceive permitted images and state.
stepExecute an action and advance physics.
finishEnd the episode and evaluate its evidence.
Observe again, track progress, and revise the plan. Observation reads do not advance physical time.
02 / A REPRESENTATIVE EVALUATION
Breadth, with a defined evaluation cost.
Meet EP-100.
A curated selection of 100 environments across five suites and 16 task types. Four prescribed conditions per task yield 400 evaluation instances at each fixed setting.
34
Launch ↗
Aim · Causal · Glass · Rules
Trajectories, indirect effects, fracture, and target constraints. 21Tower ↗
Plan · Physics · Dex
Legal rearrangements across planning depths and physical variations. 16Lock ↗
Assemble · Disassemble · Dex
Partial and complete interlocking assembly with grippers and hands. 12Cube ↗
Rubik · Pocket
Cube solving, prescribed turns, regrasping, and reorientation. 17Home ↗
Tidy · Long · Precise · Dynamic
Bookshelves, household sequences, fine manipulation, and motion.EP-Easy
The lower-workload half, with all 16 types represented.
EP-Hard
The complementary half, selected from task definitions rather than model scores.
EP-20
Ten Easy and ten Hard tasks for compact diagnosis. Pilot evidence informed selection; it is not an unbiased estimate of EP-100.
The full collection: Launcher 100 · Hanoi 120 · Pocket cube 20 · Rubik’s cube 60 · Kongming lock 60 · Bookshelf 20 · ManiSkill 30. Simulators: MuJoCo, Isaac Lab, and ManiSkill.
03 / TASK ATLAS
One loop. Different demands.
Explore the task types.
Sixteen types organize the EP-100 selection. Videos show reference executions or reproductions and their recorded outcomes, separately from the scored Codex trials.
Explore all 410 environments.
Open full catalog ↗Browse every task group and robot, with task details, task videos, and evaluation guides.
SUITE
ROBOT
No task types match this combination. Try another suite or robot.
Browse the complete EP-100 index 100 tasks →
Exact task membership, difficulty split, and EP-20 inclusion from the paper. Download index JSON ↗
04 / AGENT RESULTS
Two models. A broad test.
How far does the agent get?
Results for GPT-6-Astra and Claude · Opus-5.5-extra. The paper reports the first two valid opportunities per condition, stopping at the first success.
Performance across the five suites
Pass@1Pass@2
Headline scores give each suite equal weight. Pass@2 includes first-attempt successes; the four conditions are distinct instances, not retries.
Evaluation setting and scope How to read these results +
05 / EVALUATION PROTOCOL
Declared access. Physical success.
Every result carries its boundary.
Access levels govern eligible observations and tools, independently of task difficulty. Exact grants are declared for every run; a higher level does not automatically supply a controller.
RGB-D requires an authorized profile and environment support. Privileged observations remain read-only and preserve the same action interface and task rules.
Code-enabled control
RGB, proprioception, calibration, and public task information. Python, NumPy, Pillow, and the agent’s own algorithms.
self-written algorithmsPerception & planning
L0 plus explicitly granted reconstruction, geometry, inverse-kinematics, planning, and symbolic tools.
declared tool grantsGrasp & contact models
L1 plus explicitly granted pretrained models that propose or refine grasps and contacts.
proposal assistancePrivileged observations
L2 plus selected simulator fields exposed by a declared read-only observation profile.
state accessFour conditions per task
Prescribed variations keep the goal and rules fixed. EP-100 has 400 conditions; each gets one attempt in the base protocol, or up to two valid opportunities in the declared Codex experiment.
Loaded-only Launch
The first projectile is preloaded. Subsequent reloading remains part of multi-shot tasks. Ground-start demonstrations are outside these reported scores.
Execution, release, stability
The independent evaluator checks the executed trajectory, legal interactions, final state, and task-specific release and dwell requirements. Calling finish grants no extra settling time.
Budgets and assistance are part of the setting.
The Codex experiment allows 30M tokens and 100 images per attempt, alongside task-specific physical and wall-time limits. These are allowances, not measured usage. Pretrained action controllers and reference solutions are excluded from agent evaluation.
06 / DEMONSTRATIONS & EVIDENCE
See the interaction, know its source.
Inside the playground.
Reference executions and reproductions show the interaction through its recorded outcome. Labels identify the demonstrated task, playback speed, and evidence limits.
Home examples show L1 / Fetch ending in failure, P5 / Panda succeeding, and D1 / Panda succeeding. Each clip preserves the complete recorded sequence in at most 60 seconds; speed badges identify accelerated playback.
Sharpa cube clips demonstrate calibration, not a complete solve; Sharpa lock A06/D06 demonstrate key-only subtasks.
BUILD WITH US
The next agent is yours.
Bring a model, declare its observation and tool access, and evaluate how it turns reasoning into robotic behavior.
We will continue to expand the environments and update the demonstration videos.