Skip to content

Under maintenance. Some results may be incomplete or inaccurate.

Robotensor
All pages21

Vector · Reference

Key numbers

Every protocol constant on one page, rendered from the same spec.json the orchestrator reads.

Look-up page — built for scanning

Nothing on this site types a protocol number into its prose. Every value below is read from spec.json at build time — the same file the orchestrator reads — so what you are told here and what was actually run cannot disagree.

Changing the specification version changes every derived duel identifier and therefore every draw, so it is bumped deliberately.

The contest

The competition

Specification version
v8
Contests
1
Demonstrations in the prompt
1
Language input
none
Scored skills
1
Store schema
4

Rendered from spec.json at build time — the same file the orchestrator reads.

The contest

Contest
Vector: one demonstration from another scene, level 1
Vector: one demonstration from another scene, level 1: what the policy is shown
the frames, the action trajectory and the proprioception
Vector: one demonstration from another scene, level 1: withheld
nothing
Vector: one demonstration from another scene, level 1: where the policy starts
a different initial state
Vector: one demonstration from another scene, level 1: can the demonstration be replayed?
no — the actions do not fit a different initial state
Vector: one demonstration from another scene, level 1: scored skills
1
Vector: one demonstration from another scene, level 1: units per side, default duel
160
Vector: one demonstration from another scene, level 1: crown margin
+3.0 pts

Rendered from spec.json at build time — the same file the orchestrator reads.

The skills

In scoring order

Skill 1
Sixteen tabletop tasks (two Franka arms, level 1)
Sixteen tabletop tasks (two Franka arms, level 1) step cap
800

Rendered from spec.json at build time — the same file the orchestrator reads.

Where the units come from

Benchmarks, as distributions
robotwin: robotensor-benchmark-robotwin 0.2.0
Sixteen tabletop tasks (two Franka arms, level 1) suite
vector_level1
Sixteen tabletop tasks (two Franka arms, level 1) cameras in a clip
head_camera, left_camera, right_camera

Rendered from spec.json at build time — the same file the orchestrator reads.

What a submission is

Weights of one architecture, and nothing else

Architecture
bpp_robotwin_l1_v1
Weights file
model.safetensors
Everything a repository may hold
model.safetensors, README.md, .gitattributes
The class that serves it
vector_runtime.policy:BPPPolicy
Template the header is checked against
cfg c2dcb559ab79…, tensors 15ec2516d150…
Repository size cap
8 GiB

Rendered from spec.json at build time — the same file the orchestrator reads.

A duel

Size, at the default setting

Default duel size
launch
Selectable duel sizes
smoke, light, launch, full
Units per skill, default duel
160
Vector: one demonstration from another scene, level 1: units per side, default duel
160

Rendered from spec.json at build time — the same file the orchestrator reads.

Every size

Units per skill, smoke duel
16
Units per skill, light duel
48
Units per skill, launch duel
160
Units per skill, full duel
800

Rendered from spec.json at build time — the same file the orchestrator reads.

The crown rule

Score margin
+3.0 pts
Void units tolerated
20%

Rendered from spec.json at build time — the same file the orchestrator reads.

Budgets

Wall clocks the orchestrator enforces

Policy start budget
600 s
Per-action budget
60 s
Policy budget per unit
1,200 s
Prompt materialization wall clock
2,400 s
Per-unit wall clock
1,800 s
Per-side wall clock
12 h
Per-duel wall clock
24 h

Rendered from spec.json at build time — the same file the orchestrator reads.

Submissions

What every entry is held to

Manifest file
not this contest: submissions are weights, and nothing of them runs
Manifest api
not this contest: submissions are weights, and nothing of them runs
Policy protocol
1
Repository size cap
8 GiB
Policy base image
not this contest: submissions are weights, and nothing of them runs

Rendered from spec.json at build time — the same file the orchestrator reads.

The sandbox a policy runs in

Sandbox network
not this contest: submissions are weights, and nothing of them runs
Sandbox root filesystem
not this contest: submissions are weights, and nothing of them runs
Sandbox tmpfs size, each path
not this contest: submissions are weights, and nothing of them runs
Sandbox tmpfs executable
not this contest: submissions are weights, and nothing of them runs
Sandbox user
not this contest: submissions are weights, and nothing of them runs
Sandbox GPUs
not this contest: submissions are weights, and nothing of them runs
Sandbox CPUs
not this contest: submissions are weights, and nothing of them runs
Sandbox memory
not this contest: submissions are weights, and nothing of them runs
Sandbox process cap
not this contest: submissions are weights, and nothing of them runs

Rendered from spec.json at build time — the same file the orchestrator reads.

The sandbox is a code contest's: where submissions are code, the policy runs in it, and not this contest: submissions are weights, and nothing of them runs is its one writable path. Where they are weights, every row above says so, because nothing a participant wrote runs. What you submit says what a submission is given and asked to do.

What a policy answers with

Vector: one demonstration from another scene, level 1: action spaces a policy may declare
ee
Vector: one demonstration from another scene, level 1: action dimension
ee: 16

Rendered from spec.json at build time — the same file the orchestrator reads.

The baseline

A contest has its own genesis king, or none — a contest with no baseline opens with an empty throne and crowns its first entrant.

The genesis king

Vector: one demonstration from another scene, level 1: baseline
robotensor/vector-base

Rendered from spec.json at build time — the same file the orchestrator reads.

Clips

Every unit carries three

Clip format
mp4
Units recorded
every unit
Demonstration clip published
yes

Rendered from spec.json at build time — the same file the orchestrator reads.

Key numbers · Docs · Robotensor competitions