BWM-Forge
BWM-Forge: Multi-Task  Multi-Perspective  Multi-Scenario  Multi-Embodiment. Robot arms, quadrupeds and humanoids across eight simulated scenes.
Embodied World Programming

BWM-Forge:Agentic World Programming for Embodied Intelligence

Program geometry, motion and viewpoints. Generate high-fidelity robot videos through a collaborative agent workflow.

01 / PIPELINE · BWM-FORGE

Four agents. One editable world.

From diverse evidence to an executable scene and a high-fidelity video. Select an agent in the collaboration map to explore its internal workflow.

FROM EVIDENCE TO REPRESENTATION
First frame
maps to
Spatial scene
Prompt
maps to
Task instruction
Demonstration video
maps to
Reference demonstration
01

Space-Semantics-Driven

A reference image anchors the scene; robot assets and the task guide motion generation.

First frame + task
02

Motion Prior-Driven

Add Parquet or NPZ motion data to specify joint or end-effector motion and timing.

First frame + robot trajectory
03

Real Demonstration-Driven

Use demonstrated actions and object-state changes to reconstruct the task for the target robot.

Human demonstration video

MULTI-AGENT COLLABORATION

One shared workflow.

Four agent roles coordinate input processing, construction, review and rendering. Select an agent to explore its internal workflow.

High-fidelity task videoEditable world · review record
Solid: forward · dashed: feedback

REPROCESSING AGENT

Turn evidence into modeling context

Validate assets, align motion conventions when trajectories are present, and parse visual evidence into entities, relations and cameras.

INPUTRobot assets + task + input evidence
URDF · MEDIAAsset validation
TYPE 1Motion consistency validation
TYPE 0 · 1 · 2First-frame / demonstration parsing
MOTIONJoint mapping, units and timing
SCENE STATEEntities, relations and cameras
OUTPUTUnified modeling context
First-frame and demonstration routes share visual parsing. Type 1 adds motion consistency checks.

02 / CASES

See the world take shape.

Fourteen tasks across three input modes. Choose a result to explore its available references, modeling previews, Proxy Video and render versions.

Click a card to view its process and pause the gallery. Click elsewhere to resume.

TYPE 0 · FIRST FRAME

Chemistry experiment

FROM EVIDENCE TO RESULT

01
02
03
04

03 / EVALUATION

How video quality is evaluated

Seven objective metrics measure visual quality and motion. Three vision-language assessments check realism, coherence and task consistency.

01 / Objective metrics

Normalized to 0–100 · higher is better

018.75%

Sharpness

Clarity of image details and edges.

026.25%

Exposure

Penalizes excessively dark or bright pixels.

0312.50%

Color stability

Color consistency between adjacent frames.

048.75%

Temporal flicker

Penalizes abrupt brightness and appearance changes.

0512.50%

Motion smoothness

Continuity of optical flow and motion.

066.25%

Scene stability

Scene consistency without unwanted camera drift.

0745.00%

Action fidelity

Compare motion direction, intensity and timing against the Proxy Video.

Weights follow the current implementation.

02 / Vision-language assessment

Reference + temporal evidence + task

40%

Visual realism

Materials, lighting, shadows, geometry and generation artifacts.

25%

Physical & temporal coherence

Robot structure, contact, spatial relationships and visible continuity.

35%

Task semantic consistency

Target identity, required actions, event order and final state.

Judgments use the reference image, chronological video frames and task instructions.

Combined score · default weights

S = 0.80 Sobj + 0.20 SVLM

An initial objective score below the threshold triggers one automatic correction; videos still below the threshold proceed to human review.

04 / HARNESS & ARCHITECTURE

Every iteration leaves a useful trace.

The task coordination system (left) records execution state, artifacts and stage summaries as evidence for the experience repository (right). The repository reviews reusable rules and returns relevant experience to later Building and Rendering stages, closing the loop from execution records to experience and back to generation.

TASK COORDINATION & RECOVERY
REQUESTTask + user feedback
META-AGENTCoordinate task state
CONTEXTAssemble input and experience
FOUR AGENT ROLESExecute the selected agent stage
RECORD / RECOVERPreserve artifacts, checks and logs
STAGE SUMMARYSummarize stages and candidate rules
Task records connect input processing, construction, review and rendering. Saved state and complete intermediate outputs support recovery and execution review.
EXPERIENCE REPOSITORYEXPERIENCE CONSTRUCTION & EVOLUTION
EVIDENCEBuilding + Rendering summaries
LLM SUMMARIZATIONExtract reusable candidates
REVIEWCompare scope, rules and evidence
CREATECreate an entry
UPDATEMerge into an entry
DEACTIVATEDeactivate an entry
NO-OPLeave unchanged
EXPERIENCE REPOSITORYGeneral + specialized experience
SCOPE FILTERExclude inactive and scope-conflicting entries
TASK · ROBOT · SCENE · FEEDBACKLLM selects applicable experience
RETRIEVE → APPLY → RECORDGuide the next Building / Rendering prompt

Record. Review. Retrieve. Reuse.

Building and Rendering Agents retrieve experience before generation. Their stage documents produce new candidates; the library compares each candidate with existing rules and evidence before deciding how the repository should change.

General rules apply across tasks; specialized rules are scoped to robots, grippers, scenes, tasks or assets. A no-op leaves the repository unchanged. Retrieved rules return to generation context, completing the experience loop.

01 / COORDINATE

Control the next step

Keep the current stage, iteration and feedback aligned as agents exchange results.

02 / PRESERVE

Keep the evidence

Retain project previews, execution logs, checks, video versions and user decisions.

03 / RECOVER

Continue from recorded state

Recover the task context and intermediate results to review or continue an interrupted workflow.

Generate synchronized original, following third-person and approximate first-person views from one scene. Edit camera paths on a shared motion timeline.

Program object relationships and motion for manipulation and locomotion tasks. Use feedback to refine the relevant part of the scene program.

Use one workflow across single arms, dual arms with grippers or dexterous hands, quadrupeds and humanoids.

Render one Proxy Video in different scene styles. Reference images and prompts vary appearance while the proxy guides motion and timing.