π0.5
Physical Intelligence
78.4ILLUSTRATIVE INDEX
ROBOT RANKINGS / EVIDENCE MENUModels meet
Models meet
machines.
Explore robot intelligence, embodiments, edge compute, benchmarks, realtime behavior, and long-horizon deployment evidence in one index.
Important: Jetson is an edge-compute platform—not a robot model. Embodied Arena keeps model, runtime, hardware, and embodiment separate.Intelligence with an embodiment.
A model score is meaningful only when the robot body, sensors, compute, task distribution, and intervention policy travel with it.
GR00T N1.6
NVIDIA
75.9ILLUSTRATIVE INDEX
OpenVLA-OFT
Stanford / TRI
72.6ILLUSTRATIVE INDEX
Rank with context.
Filter the public-beta seed set. Scores remain illustrative until submissions share a frozen protocol and comparable evidence.
π0.5OPENPhysical Intelligence · GENERALIST VLA
78.4
● REALGR00T N1.6OPENNVIDIA · HUMANOID VLA
75.9
● REALOpenVLA-OFTOPENStanford / TRI · OPEN VLA
72.6
● REALHelix 02Figure AI · HUMANOID POLICY
69.8
● REALRoboBrain 2.0OPENBAAI · EMBODIED VLM
67.2
● SIMSmolVLAOPENHugging Face · COMPACT VLA
58.7
● REALIllustrative public-beta data. Hardware, tasks, training data, sample sizes, and evaluation protocols differ; do not treat this table as a deployment claim.
The winner changes with the work.
Compare within matched task families rather than collapsing every form of physical intelligence into one number.
π0.5
86.2Seed normalized index
GR00T N1.6
80.4VLA + whole-body policy
RoboBrain 2.0
81.2Simulation evidence
OpenVLA-OFT
79.3Open weights
Price the useful outcome.
Robot economics must include compute, energy, human attention, retries, maintenance, and downtime—not API cost alone.
Cmission=
compute + energy + operator time + recovery + maintenance
No comparable cost ranking is published yet. Submissions must disclose every term before this section ranks systems.
Four layers, disclosed separately.
Embodied Arena never credits a compute module with the capabilities of the policy running on it—or a model with the safety properties of the robot around it.
Perception, reasoning, and action policy.
On-robot inference and sensor processing.
Drivers, acceleration, middleware, and tools.
Body, actuators, sensors, and safety envelope.
Choose the claim before the test.
Short task success, realtime interaction, and long-horizon autonomy answer different questions. Use the smallest protocol that supports the intended claim.
01
HILO REALTIME
↗02Human burden, MTHI, interruption, latency, and safety-kernel evidence.
WANTED-10K
↗03Long-horizon resident hours, intervention burden, reliability, and evidence integrity.
LIBERO / CALVIN
↗04Short-horizon manipulation task suites. Useful, but not field autonomy certificates.
REAL OPERATIONS
↗Throughput, recovery, service continuity, and operator burden under deployment load.
Latency is a chain.
Measure camera exposure to perception, reasoning, safety decision, command dispatch, and physical response as separate p50, p95, and p99 stages.
01IMAGE EVENT02PERCEPTION03POLICY04SAFETY KERNEL05ACTUATION
No cross-platform speed winner is declared without matched power mode, sensor rate, quantization, batch size, and robot workload.
Bodies are part of the benchmark.
A policy transfer across arms, mobile manipulators, humanoids, quadrupeds, and autonomous mobile robots is evidence—not an implementation detail.
01HUMANOID02DUAL ARM03MOBILE MANIPULATOR04ROBOT ARM05QUADRUPED06AMR
Jetson, correctly placed.
NVIDIA Jetson modules provide the on-robot compute layer. Rank them on matched robotics workloads, sustained latency, power, thermals, memory headroom, and sensor concurrency.
Jetson AGX Thor
Up to 2,070 FP4 TFLOPS- MEMORY
- Up to 128 GB
- POWER
- 40–130 W
- BEST FIT
- Humanoids + physical AI
Jetson AGX Orin
Up to 275 TOPS- MEMORY
- 32 / 64 GB
- POWER
- Configurable
- BEST FIT
- Multi-sensor autonomous machines
Jetson Orin NX
Up to 157 TOPS- MEMORY
- 8 / 16 GB
- POWER
- Compact
- BEST FIT
- Mobile robots + manipulators
Jetson Orin Nano
Up to 67 TOPS- MEMORY
- 4 / 8 GB
- POWER
- 7–25 W
- BEST FIT
- Entry edge AI + prototypes
Robots need more than tokens.
Disclose the operational memory that actually affects behavior: observation window, map lifetime, session continuity, skill state, and human instruction history.
Tool calls become physical events.
Every proposed action should bind to a safety decision, robot command, result, and recovery path in the HILO event chain.
01observe()02plan()03navigate()04move()05grasp()06inspect()07ask_human()08yield()09recover()10stop()Cameras become image events.
Realtime models consume selected, timestamped image events—not an undocumented firehose. Preserve capture time, selection policy, redaction, inference binding, and dropped-frame evidence.
01CAMERA→02LOCAL REDACTION→03FRAME SELECTOR→04IMAGE EVENT→05MODEL→06ACTION
The long horizon wins.
Promote systems by independently verified resident exposure—not polished demos or isolated successful episodes.
Information architecture adapted from the OpenRouter Rankings section menu. OpenRouter rankings data is not reproduced. Reference data is licensed under CC BY 4.0; robotics labels, taxonomy, and evidence rules are Embodied Arena adaptations.