XEbench
Measure · cross-embodiment
Same task on many real robots.
XEsim
Replicate · cross-environment
Every real run rebuilt in sim.
Robot-RL
Train · same tools, same verifier
Verifier success becomes the reward.
5
embodiments, more coming
12
real tasks
2×
real + replica
400
sim tasks for RL
Built together by
What XE means
Cross-embodiment × cross-environment
Same task across bodies. Same run in real and in sim.
YAM pair
xArm pair
xArm + Wuji Hand 2
Humanoid
Mobile manipulator
Real
Recording
Recording
Recording
Coming
Coming
Sim
First replicas
First replicas
First replicas
Coming
Coming
Real vs sim
Every recording gets a replica
Drag to compare the real robot with its replica replaying the same joint motion.
realreplica
YAM pair, 3× speed. First MuJoCo replicas, before the contact work.
One storyline
↺ trained agents return to XEbench
XEbenchcross-embodiment · same task, many bodies
Robots
Same task, many bodies
YAM pair
Two 6-DoF arms · parallel grippers
xArm pair
Two arms · parallel grippers
xArm + Wuji Hand 2
Five-finger dexterous hand
More bodies coming next
Humanoid · mobile manipulator · more hands
Tasks
Twelve tasks, from atomic to contact-rich
Three tiers: anchor, middle, frontier.
parallel handsanchor
Sort lollipops and snacks into a plate
parallel handsmiddle
Stack the two outer cups into the middle one
one holds, one actsanchor
Pull the cap off a marker
one holds, one actsmiddle
Uncap and recap a marker
handoveranchor
Hand an object across, then into a box
one holds, one actsmiddle
Hold a test-tube rack and insert a tube
one holds, one actsfrontier
Hold a network switch and plug in an Ethernet cable
soft object, two handsanchor
Fold a T-shirt or towel in half
soft object, two handsmiddle
Fold a towel twice
soft object, two handsfrontier
Fold a T-shirt
rigid object, two handsanchor
Carry an empty tray, keeping it level
rigid object, two handsmiddle
Carry a tray with a cup on it
More tasks coming
plannedHarder
- Unscrew a jar lid
- Pour into a cup without spilling
- Route a cable through clips
- Fold a T-shirt and stack it
Several robots together
- A humanoid and a mobile manipulator tidy a room together
- Two robots carry a long board
- Pass an object between stations
- Assembly line: open, pack, close a box
Inside the lab
Same brain, different bodies
Four models, three real arms, one task: put a brick in a box.
XEsimcross-environment · every recording gets a replica
How a replica is built
Robot logs
Controller parameters
Video
Camera calibration
Real scene
Object poses, background
Replay + compare
MuJoCo or IPC contact
A replica is useful when it ranks agents in the same order as the real robot. See the comparison ↑
Run protocol
Same prompt, same tools, same verifier
Only the backend changes between a real run and its replica.
{
"name": "move_to",
"arguments": {
"observation_id": "obs_0042",
"command_id": "cmd_0043",
"note": "lift the cup 5 cm",
"frame": "base",
"translation_m": [0.0, 0.0, 0.05],
"rotation_delta_rad": [0.0, 0.0, 0.0]
}
}agent: cc # cc · codex · dsh profile: intuitive # or expert engine: kind: robosuite # replica in sim # kind: real # the real station task: cup_stack command_budget: 80
$ ./scripts/run.sh demo \ --agent cc \ --claude .tools/claude-2.1.296/claude \ --provider openai # → "demo_verified": true
Agent harness
Claude Code · Codex · DeepSeek Harness
Robot tools over MCP
intuitive: 5 tools · expert: 16
Real robot or replica
Same tool calls, either backend
Separate verifier
Scores the final state
Robot-RLtrain · same tools, same verifier
The arena is also a training environment
The verifier that scores a run is the reward that trains the agent.
replica · 12×
Next: XEsim replicas as training environments.
5 / 16tools: intuitive / expert
4engines: MuJoCo · Newton · OvPhysX · real
400Tabletop400 training tasks
27BQwen3.8 GRPO recipe
$ bash examples/train_qwen38_27b.sh --dry-run \ --actor-num-nodes 2 --actor-num-gpus-per-node 4 \ --rollout-num-gpus 4 \ --tensor-model-parallel-size 2 --pipeline-model-parallel-size 2 \ --rollout-batch-size 8 --n-samples-per-prompt 8 # review the generated command, then drop --dry-run
Runs in simulation today. A full 27B training step is not yet verified.
Status
Where things stand
real
Robot runs
Recordings on YAM, xArm and Wuji Hand 2.
replica
Contact physics
IPC holds the cups; cloth close, not matched yet.
benchmark
Scores
No paired real and sim scores yet.
training
Robot-RL
Runs in sim; full 27B step pending.
Related benchmarks
What is new here
| Benchmark | Tests | Cross-embodiment | Cross-environment |
|---|---|---|---|
| Embodied First Exam | coding agents | yes | sim only |
| Ember-0 (XDOF) | frontier LLMs | one robot type | real only |
| RoboDojo | learned policies | tasks differ per platform | separate task sets |
| DuoBench | learned policies | one robot | 4 of 11 tasks |
| SO-101 Bench | learned policies, one LLM agent | one arm | yes, Isaac Lab twin |
| LIBERO | learned policies | one arm (Franka) | sim only |
| RoboCasa | learned policies | several robots, one sim | sim only |
| RoboTwin 2.0 | learned policies | 5 dual-arm robots | sim only |
| BEHAVIOR-1K | learned policies | a few mobile manipulators | sim only |
| XE2 Arena | LLM agents + RL training | five embodiments, more coming | a replica per recording |
Contact
Run your model on XE2 Arena
Want your agent or policy on our robots and replicas, or want to collaborate? Talk to our researchers.