Cyborg Index · July 2026
Aster-02 vs Halide-03
Both humanoids read on the same 8 axes and 40 sub-axes, using the published Index scores. Every difference below carries the public evidence behind it, and sub-axes with no public evidence count as 0 rather than being skipped.
Overall standing
Capability score · 0–5 over 40 sub-axesThe short answer
Aster-02 and Halide-03 score the same — 3.34 and 3.35.
Nothing separates them overall. The table below shows the axes where they actually differ.
Where Aster-02 is ahead
- Deployment+2.16Independently backed
- HRI+0.58Independently backed
- Locomotion+0.37Independently backed
- Safety+0.07Independently backed
Where Halide-03 is ahead
- Learning+1.13Rests on claims and demos
- Autonomy+0.92Rests on claims and demos
- Manipulation+0.72Independently backed
- Perception+0.49Independently backed
What would change it
7 sub-axes on the trailing side have no public evidence yet. If all of them scored full marks they would be worth 0.88 — against a gap of 0.01.
- Manipulation · In hand reorientup to +0.125
- Locomotion · Disturbance recoveryup to +0.125
- Learning · Sim to realup to +0.125
- Locomotion · Stairsup to +0.125
- Autonomy · Unattended operationup to +0.125
How much we know
91%of sub-axes covered
We have public evidence for 91% of the 40 sub-axes across these two.
Aster-02 · 83%
Halide-03 · 100%
The shape of each score
33/40 sub-axes have evidence
Faint outline = Halide-03
40/40 sub-axes have evidence
Faint outline = Aster-02
Reading the wheels — clockwise from the top
- 1 Locomotion
- 2 Manipulation
- 3 Perception
- 4 Autonomy
- 5 Learning
- 6 HRI
- 7 Deployment
- 8 Safety
Each wheel has 40 spokes — five per axis. A longer spoke is a higher score, and the dashed rings mark 1 to 4. Hover or tab onto any spoke to read the axis, the sub-axis and the exact value. A missing spoke means there is no public evidence for that sub-axis yet, which is not the same as scoring zero. The faint outline on each wheel is the other humanoid, drawn on the same scale.
Every axis, every sub-axis
Aster-02Halide-03
Shared 0–5 track · widest gap firstDeployment
4.22 / 2.06◄ 2.16
Independently backedEvidenced cells · 4/5 vs 5/5
Named pilots
4.44 / 2.48◄ 1.96
Field hours
4.40 / 2.30◄ 2.10
Customer diversity
4.13 / 2.25◄ 1.88
Supportability
4.22 / 1.70◄ 2.52
Commercial posture
3.92 / 1.56◄ 2.36
Learning
2.95 / 4.09► 1.13
Rests on claims and demosEvidenced cells · 4/5 vs 5/5
Imitation learning
2.80 / 4.19► 1.39
Sim to real
3.31 / 4.14► 0.83
Foundation policy
3.24 / 3.79► 0.55
Data flywheel
2.73 / 4.17► 1.44
Transfer fine tune
2.69 / 4.15► 1.46
Autonomy
2.58 / 3.50► 0.92
Rests on claims and demosEvidenced cells · 4/5 vs 5/5
Unattended operation
2.30 / 3.17► 0.87
Task horizon
2.87 / 3.76► 0.89
Generalization
2.92 / 3.61► 0.69
Instruction following
2.11 / 3.93► 1.82
Recovery from failure
2.68 / 3.03► 0.35
Manipulation
3.36 / 4.08► 0.72
Independently backedEvidenced cells · 4/5 vs 5/5
Grasp variety
3.27 / 4.07► 0.80
Bimanual
3.07 / 3.97► 0.90
Force control
3.06 / 3.81► 0.75
In hand reorient
3.47 / 4.46► 0.99
Tool use
3.92 / 4.07► 0.15
HRI
3.19 / 2.61◄ 0.58
Independently backedEvidenced cells · 4/5 vs 5/5
Speech io
3.45 / 2.46◄ 0.99
Nl instruction
3.37 / 2.44◄ 0.93
Intent recognition
3.31 / 2.53◄ 0.78
Safe proximity
3.18 / 2.97◄ 0.21
Social presentation
2.65 / 2.66
Perception
3.38 / 3.87► 0.49
Independently backedEvidenced cells · 5/5 vs 5/5
Sensor stack
3.23 / 4.14► 0.91
Scene understanding
3.72 / 3.56◄ 0.16
Spatial memory
3.46 / 3.99► 0.53
Object permanence
3.58 / 4.01► 0.43
Semantic world model
2.92 / 3.66► 0.74
Locomotion
4.00 / 3.63◄ 0.37
Independently backedEvidenced cells · 3/5 vs 5/5
Flat ground walking
4.26 / 3.11◄ 1.15
Uneven terrain
4.11 / 3.69◄ 0.42
Stairs
3.83 / 3.51◄ 0.32
Disturbance recovery
4.11 / 4.08
Speed endurance
3.70 / 3.78► 0.08
Safety
3.06 / 2.99◄ 0.07
Independently backedEvidenced cells · 5/5 vs 5/5
Functional safety
2.79 / 2.68◄ 0.11
Certifications
2.90 / 2.88
E stop arch
2.87 / 3.06► 0.19
Incident transparency
3.40 / 3.16◄ 0.24
Force impact limits
3.34 / 3.19◄ 0.15
Segment length and thickness are the gapNo public evidence yet
Which is better for what
Each row is decided by the axes that matter for that kind of work, read off the same Capability scores as the ranking. Rows only appear where both humanoids have enough public record to answer.
Manufacturing and warehouse work
Named pilots and field hours, backed by what the hands can actually do.
Aster-02
Aster-02 scores higher on Deployment and Manipulation — 3.93 against 2.73 for Halide-03.
Running without a human in the loop
Unattended operation and task horizon, backed by what the robot can perceive.
Halide-03
Halide-03 scores higher on Autonomy and Perception — 3.62 against 2.84 for Aster-02.
Moving through difficult ground
Uneven terrain, stairs and disturbance recovery.
Aster-02
Aster-02 scores higher on Locomotion — 4.00 against 3.63 for Halide-03.
Working alongside people
Functional safety and certification, plus how it reads and answers people.
Aster-02
Aster-02 scores higher on Safety and HRI — 3.10 against 2.87 for Halide-03.
Picking up new tasks
Imitation learning, transfer and generalisation across tasks.
Halide-03
Halide-03 scores higher on Learning and Autonomy — 3.89 against 2.83 for Aster-02.
Research and developer use
How much of the stack is documented and reproducible.
Halide-03
Halide-03 scores higher on Learning and Perception — 4.02 against 3.10 for Aster-02.
How solid the record is on each side
Aster-02
- Sub-axes on the record
- 33 of 40 · 83%
- Carried by verified material
- 74%
- Rests on claims and demos
- 0.88
Best documented — Locomotion, Manipulation, Perception, Autonomy, Learning, HRI, Deployment, Safety
Nothing on the record — every axis has at least one read
Halide-03
- Sub-axes on the record
- 40 of 40 · 100%
- Carried by verified material
- 58%
- Rests on claims and demos
- 1.40
Best documented — Locomotion, Manipulation, Perception, Autonomy, Learning, HRI, Deployment, Safety
Nothing on the record — every axis has at least one read
What this comparison cannot settle
- 7 sub-axes are read on only one side. A one-sided cell is undecided, not a zero for the silent side.
- Demonstrations happen in different environments and product revisions are not always equivalent, so like-for-like uptime and throughput figures do not exist for most of the field.