M9G | Technical Research Blog
M9G-TR-2026-06 • Browser-Native 3D Multi-Robot TAMP & Smart Factory Engineering Published November 9, 2026 • Live WebGL Deployment: https://tamp.mincasurong.com ↗

Engineering tamp.mincasurong.com: Zero-Install 3D Multi-Robot TAMP and a 6-Cell Smart Factory in 60 FPS WebGL

Testing multi-arm Task and Motion Planning (TAMP), pre-dispatch safety gates, and fleet scheduling normally requires a 24 GB VRAM workstation running NVIDIA Isaac Sim or a ROS 2 and MoveIt 2 container cluster. To make our Lazy-DaSH and Project STAMP stacks runnable in any browser, we built and deployed tamp.mincasurong.com (fast_sim_web). This client-side 60 FPS Three.js WebGL simulator pairs an 8-slot Franka FR3 workcell with a 6-cell, 18-FR3, and 4-AMR smart factory at 720.0 units/h. Below are the five bottlenecks that block real-time 3D TAMP in the browser and the exact mesh, kinematic, scheduling, clearance, and layout mechanisms that solve them.

M9G-TR-2026-06 • Browser-Native 3D Multi-Robot TAMP & 6-Cell Smart Factory Lite-Sim at tamp.mincasurong.com
TL;DR: Five Browser TAMP Bottlenecks Solved at tamp.mincasurong.com ↗
  • 1. 134 KB Binary STL Pack Instead of a 24 GB VRAM Workstation Boot: Rather than downloading multi-gigabyte USD scenes or streaming laggy cloud WebRTC video, tamp.mincasurong.com packs decimated binary STL link meshes (link0..7.stl and hand.stl, 134 KB total), clones Three.js BufferGeometry across up to 22 FR3 arms, boots in under 400 ms, and holds 60 FPS on laptops and mobile browsers.
  • 2. Sub-25 ms TypeScript 7-DoF DLS IK and 128 Hz Swept-Capsule Gate: fr3Kinematics.ts solves 150-iteration 7-DoF DLS IK in under 25 ms using Craig Modified Denavit-Hartenberg Forward Kinematics and an analytical 6×7 Geometric Jacobian (damping 0.02, null-space posture weight 0.05). A 128 Hz swept-capsule gate checks every trajectory before execution and spawns a translucent red 3D ghost arm at any rejected pose.
  • 3. Sub-60 ms 8-Slot (s0..s7) Lazy-DaSH Replanning Across Three Modes: Toggling any of the 8 peripheral workcell slots triggers TypeScript Lazy-DaSH replanning in 27.1 to 34.7 ms across Mode 1 (6-Block Tower: 9.52 s at 2 arms → 7.97 s at 3 arms → 6.56 s at 4 arms in 30.8 ms), Mode 2 (Bimanual LongBar & Conveyor: 4 arms in 5.37 s with 2 conflict repairs in 27.1 ms), and Mode 3 (Deformable Cloth: catching 2-arm 1_IK_UNREACHABLE at the 0.910 m pin beyond 0.805 m reach and recovering in 34.7 ms at 7.91 s once slot s6 is enabled).
  • 4. Elevated Z = +0.02 m AMR Turret Docking, Z = -0.52 m Floor Clearance, and ResizeObserver Hypergraph UI: Across the 6-cell, 18-FR3, and 4-AMR (MM0..MM3) smart factory running at 720.0 units/h, depressing the floor to Z = -0.52 m and raising each AMR turret base to Z = +0.02 m eliminates Z-fighting and table-apron collisions, while pointer-captured split resizers and a ResizeObserver scale the Execution Hypergraph and Gantt timeline without label clipping.
FR3 Mesh Payload
134 KB Binary STL
link0..7.stl + hand.stl at 60 FPS WebGL
7-DoF DLS IK & Gate
< 25 ms · 128 Hz
fr3Kinematics.ts + Swept Capsule Gate
Lazy-DaSH Replan
27.1 – 34.7 ms
8-Slot Workcell (Tower · LongBar · Cloth)
6-Cell Smart Factory
720.0 units/h
18 FR3 + 4 AMRs (60.0 s · 12 orders)
TL;DR Video Synthesis: 40s Walkthrough of Browser-Native 3D TAMP and the 6-Cell Smart Factory at tamp.mincasurong.com ↗ Direct MP4 Stream (blog6-wrapup-omni.mp4) ↗
Stage 1 • 134 KB Binary STL & 7-DoF MDH FK
Sub-400 ms 60 FPS Three.js Kinematic Scene

Fetches link0..7.stl and hand.stl (134 KB total) once and clones BufferGeometry across 8 workcell slots or 22 factory FR3 arms.

Stage 2 • 6×7 DLS IK & 128 Hz Capsule Gate
fr3Kinematics.ts (<25 ms) + 3D Ghost Inspector

Solves 150-iteration Gaussian Elimination DLS IK (damping 0.02, posture weight 0.05) and checks 128 Hz swept link capsules before motion.

Stage 3 • Lazy-DaSH Replanner (27.1–34.7 ms)
8-Slot Conflict Repair & Slot s6 Recovery

Repairs frontier collisions, bimanual promotions, and unreachable cloth pins well under the 60 ms interactive budget across Modes 1–3.

Stage 4 • 6-Cell Smart Factory & Split Dock
18 FR3 + 4 AMRs at 720.0 units/h

Docks MM0..MM3 turrets at Z = +0.02 m over a Z = -0.52 m floor while scaling Hypergraph and Gantt panes via ResizeObserver.

1. Replacing a 24 GB VRAM Workstation Boot with a 134 KB Binary STL Pack at 60 FPS

Desktop robotics simulators block instant browser verification. Our workstation stack in IsaacSim_Gemini runs NVIDIA Isaac Sim 6.0 (Omniverse Kit and 240 Hz PhysX 5 tensors) alongside WSL2 ROS 2 Jazzy, MoveIt 2, and MuJoCo 3.x. That setup requires an RTX 3090 or 4090 GPU with 16 to 24 GB of VRAM, 45 to 90 seconds of cold boot time, and more than 4.5 GB of USD assets. Reviewers on standard laptops or tablets cannot run it locally.

Server-side Omniverse WebRTC pixel streaming moves that GPU bill to the cloud while adding session queues, video compression blur, and round-trip input lag on every camera orbit or slot click. Instead, https://tamp.mincasurong.com runs fast_sim_web 100% client-side in WebGL at a locked 60 FPS:

Client-side architecture of fast_sim_web at tamp.mincasurong.com: 134 KB Binary STL WebGL Scene, fr3Kinematics.ts 7-DoF MDH FK and 6x7 DLS IK Solver, 128 Hz Swept-Capsule Gate, Lazy-DaSH Replanner, and 6-Cell Smart Factory
Figure 1 (M9G-TR-2026-06). Client-side architecture of fast_sim_web at https://tamp.mincasurong.com: 134 KB binary STL mesh cache (link0..7.stl and hand.stl), TypeScript 7-DoF MDH FK and sub-25 ms 6x7 DLS IK in fr3Kinematics.ts, 128 Hz swept-capsule safety gate with 3D red collision ghosts, 27.1–34.7 ms Lazy-DaSH replanner, and the 6-cell Smart Factory orchestrator.

2. Single-Threaded 7-DoF MDH Kinematics, Sub-25 ms DLS IK, and the 128 Hz Swept-Capsule Gate

Browsers run UI and WebGL rendering on one main thread without native C++ MoveIt 2, KDL, or CUDA cuRobo bindings. If an inverse-kinematics solver stalls for 100 ms or allocates thousands of temporary matrices per frame, garbage-collection pauses drop the viewport below 60 FPS. Conversely, simplified 3-DoF planar approximations miss 7-DoF elbow redundancy, wrist singularities, and URDF joint-limit saturations on the Franka FR3.

In tamp.mincasurong.com, fr3Kinematics.ts solves 150-iteration 7-DoF DLS IK in under 25 ms with pre-allocated typed arrays and pairs directly with a 128 Hz swept-capsule collision gate:

Browser-Native Franka FR3 7-DoF Kinematic & Swept-Capsule Gate Specification Ref: M9G-TR-2026-06 • fast_sim_web/src/fr3Kinematics.ts
1. 7-DoF Modified Denavit-Hartenberg (MDH) Forward Kinematics: Chains Craig Modified DH transforms (d1 = 0.333 m, d3 = 0.316 m, d5 = 0.384 m, a4 = +0.0825 m, a5 = -0.0825 m, a7 = 0.088 m, flange offset 0.107 m) to compute all seven joint frames and the end-effector tool center point (TCP).
2. Analytical 6×7 Geometric Jacobian: Reads each revolute joint's 3D axis and origin from the forward chain to fill the 3×7 linear velocity rows (cross product of joint axis and TCP lever arm) and 3×7 angular velocity rows without finite-difference jitter.
3. Gaussian Elimination Damped Least-Squares (DLS) IK: fr3Kinematics.ts solves the 6×6 damped normal system (lambda = 0.02) with a null-space posture term (weight 0.05 toward upright home), converging to sub-millimeter TCP error within 150 iterations in under 25 ms while clamping all seven joints to FR3 URDF bounds.
4. 128 Hz Swept-Capsule Safety Gate: Samples candidate multi-arm trajectories at 128 Hz, computes analytical capsule-to-capsule and capsule-to-AABB clearances across all links, emits deterministic fault codes (IK_UNREACHABLE, WORKSPACE_COLLISION, BIMANUAL_PROMOTION_REQUIRED), and spawns a translucent red 3D ghost manipulator at the exact rejected pose.

3. Sub-60 ms 8-Slot (s0..s7) Lazy-DaSH Replanning Across Tower, LongBar, and Deformable Cloth

In the interactive workcell at tamp.mincasurong.com, users toggle any subset of 8 peripheral mounting slots (s0..s7), drop workspace obstacles, and switch across three manipulation modes. Rebuilding a coupled motion graph from scratch on every click would stall the browser.

Running UIUC Parasol Lab's Lazy-DaSH (Lee et al., IEEE T-RO) task-space hypergraph search and branch-scoped conflict repair in TypeScript keeps full multi-arm replanning between 27.1 ms and 34.7 ms, staying well inside our 60 ms interactive budget:

Table 1 • Verified Browser Telemetry Across the 3 Cooperative Workcell Modes at tamp.mincasurong.com (Slots s0..s7) M9G-TR-2026-06 • Live Browser Benchmark
Workcell Mode Active Slot & Arm Topology Physical & Kinematic Constraint Tested Verified Makespan & Lazy-DaSH Replan Time
Mode 1: 6-Block Cooperative Tower Scales live across 2 Arms → 3 Arms → 4 Arms on slots s0..s7 Strict vertical stacking order at the central tower goal with spatio-temporal mutex locking across concurrent outer-table picks and central placements. Makespan drops from 9.52 s (2 arms) → 7.97 s (3 arms) → 6.56 s (4 arms); full hypergraph schedule and conflict check finishes in 30.8 ms.
Mode 2: Bimanual LongBar & Moving Conveyor 4 Arms cooperative cell with live belt pickup Emits BIMANUAL_PROMOTION_REQUIRED if one arm attempts a solo lift on the high-moment LongBar; locks dual-arm grasp keyframes while clearing belt blockers. Completes synchronized dual-arm lift and blocker clearance in 5.37 s makespan, resolving 2 conflict repairs in 27.1 ms.
Mode 3: Deformable Cloth Unfold, Fold & Handover 2 Arms (Halt) → 3+ Arms with Slot s6 Enabled Pin-and-pull unfolding requires holding a far cloth corner at 0.910 m radius, exceeding the 2-arm kinematic reach limit of 0.805 m. 2-arm Gate halts with 1_IK_UNREACHABLE (0.910 m > 0.805 m); enabling slot s6 replans in 34.7 ms and completes the fold and handover in 7.91 s.
Figure 2A. Workcell Makespan Scaling & 27.1–34.7 ms Lazy-DaSH Replanning (Modes 1–3) tamp.mincasurong.com • Slots s0..s7
Mode 1 (6-Block Tower) · 2 Active FR3 Arms9.52 s Makespan (Baseline 2-Arm Serialized Handover)
Mode 1 (6-Block Tower) · 3 Active FR3 Arms7.97 s Makespan (-16.3%) (3-Way Pipelined Pick & Place)
Mode 1 (6-Block Tower) · 4 Active FR3 Arms (30.8 ms Replan)6.56 s Makespan (-31.1%) (4-Arm Parallel Wavefront)
Mode 2 (Bimanual LongBar & Conveyor) · 4 Arms (27.1 ms Replan, 2 Repairs)5.37 s Makespan (Synchronized Dual-Arm Lift)
Mode 3 (Deformable Cloth) · 2 Arms (0.910 m Pin > 0.805 m Max Reach)HALT: 1_IK_UNREACHABLE (Blocked Pre-Dispatch → Enable Slot s6)
Mode 3 (Deformable Cloth) · Slot s6 Enabled (34.7 ms Recovery Replan)7.91 s Makespan (100% Pin-and-Pull & Handover Completion)

4. Smart Factory Clearance: Elevated Z = +0.02 m AMR Turrets and Z = -0.52 m Floor Datum at 720.0 units/h

Scaling tamp.mincasurong.com from one 8-slot table to a 6-Cell Smart Factory with 18 cell-mounted FR3 arms (3 per cell across 6 cells), 4 autonomous mobile manipulators (MM0..MM3), 3-tier AS/RS storage racks, and a roller shipping dock exposed three geometric failures at once.

In early builds, co-planar floor and pedestal meshes caused depth-buffer Z-fighting, low-slung AMR arms clipped table aprons with Link 1 and Link 2 during reach-in, and crowded racks blocked aisle sightlines. Five geometric invariants eliminated all three issues and locked throughput at 720.0 units/h:

6-Cell Smart Factory Spatial Clearance & Elevation Invariants Ref: M9G-TR-2026-06 • fast_sim_web Smart Factory Mode
1. Floor Plane Depression (Z = -0.52 m): Sets the global factory floor and AMR travel aisles at Z = -0.52 m relative to the workcell tabletop datum (Z = 0.00 m), ending depth-buffer Z-fighting between AMR wheels, cell pedestals, and table legs.
2. Elevated AMR Turret Riser (Z = +0.02 m Base Flange): Mounts a vertical turret column on each mobile base (MM0..MM3) so the FR3 link0 flange sits at Z = +0.02 m above the tabletop, matching the downward reach-in clearance of table-mounted arms.
3. Raised Overhead Transfer Waypoints: Lifts transit and handover Z-waypoints during AMR-to-cell and cell-to-shipping-dock transfers so swept forearm and gripper capsules clear tabletop fixtures.
4. Interior Table-Grid Clamping: Clamps all buffer slots, WIP trays, and assembly targets inside the interior table polygon, keeping grasps away from bevelled table edges.
5. 50% AS/RS Rack & AMR Fleet Right-Sizing: Cuts rack footprint and mobile manipulator count by 50% (streamlining to MM0..MM3 and compact 3-tier racks), clearing aisle sightlines while sustaining 720.0 units/h (60.0 s makespan across 12 multi-stage orders).
Figure 2B. Browser Frame Budget & 6-Cell Smart Factory Throughput (720.0 units/h) 18 Cell FR3 + 4 Mobile Manipulators (MM0..MM3)
7-DoF DLS IK Solve in fr3Kinematics.ts (150 Max Iterations, Damping 0.02)< 25.0 ms (41.7% of 60 ms Interactive Replan Budget)
Lazy-DaSH Workcell Schedule & Conflict Repair (Modes 1–3)27.1 – 34.7 ms (Sub-60 ms Single-Thread Browser Target)
6-Cell Smart Factory Fleet Throughput & Order Completion (12 Orders / 60.0 s)720.0 units/h (100% Order Completion) (0 AMR or Table Collisions)

5. Split-Pane Telemetry Dock and ResizeObserver Execution Hypergraph Scaling

Inspecting a multi-arm planner requires more than a 3D camera. Alongside the WebGL scene, tamp.mincasurong.com streams the Execution Hypergraph (task vertices, active composite hyperarcs, and branch-scoped repairs), the Multi-Arm Spatio-Temporal Gantt Timeline, and the Deterministic Safety Gate Log in real time.

Fixed CSS grids clip SVG node labels and collide timeline bars when a 4-to-22 robot Gantt chart shares a laptop screen with a 3D viewport. Two layout mechanisms keep every pane legible during live 60 FPS playback:

  1. Pointer-Captured Split Resizers (.resizer-h, .resizer-v-inner, .resizer-v): Drag handles between the 3D WebGL canvas, control sidebar, Execution Hypergraph pane, and bottom Gantt dock capture pointer events and clamp flex ratios within safe minimum and maximum bounds.
  2. ResizeObserver Node and Typography Scaling: A dedicated ResizeObserver watches the pixel width and height of the Execution Hypergraph and Gantt containers on every drag frame, scaling SVG node radii, edge stroke widths, and clamp() font sizes proportionally so state labels and timeline bars never clip or overlap.
Table 2 • Systems Comparison: Workstation Isaac Sim & ROS 2 vs. Cloud WebRTC vs. tamp.mincasurong.com (fast_sim_web) M9G-TR-2026-06 Systems Matrix
System Dimension Workstation Isaac Sim 6.0 + ROS 2 / MoveIt 2 Cloud GPU WebRTC Streaming tamp.mincasurong.com (fast_sim_web Lite-Sim)
Hardware & VRAM Requirement NVIDIA RTX 3090 or 4090 (16–24 GB VRAM) + 64 GB RAM + Ubuntu / WSL2 Dedicated cloud GPU node ($1.50–$4.00/h per active viewer) 0 Install · Any Laptop, Tablet, or Mobile Browser (Standard WebGL)
Asset Payload & Cold Boot 4.5+ GB USD & URDF assets · 45–90 s simulator boot 15–30 s container spin-up + WebRTC handshake delay 134 KB Binary STL Pack (link0..7.stl, hand.stl) · < 400 ms Boot
Kinematics & Safety Gate C++ MoveIt 2 / CUDA cuRobo + Python stamp_gate Server-bound ROS 2 node (network round-trip on every drag) TypeScript 7-DoF MDH FK + 6×7 DLS IK (<25 ms in fr3Kinematics.ts) + 128 Hz Capsule Gate
Workcell & Factory Modes Separate Python launch files per workcell or MuJoCo batch Single fixed scene per stream session 8-Slot Workcell (Modes 1–3) + 6-Cell Smart Factory (18 FR3 + 4 AMRs at 720.0 units/h)
Telemetry & Hypergraph UI External RViz2, Foxglove, or matplotlib windows Baked into compressed video stream (non-interactive text) Draggable Split Dock (.resizer-h, .resizer-v-inner) + ResizeObserver Hypergraph & Gantt

6. References, Live Deployment & Related M9G Reports

  1. Live Production Simulator (fast_sim_web): Minsu Chang, Ph.D., "M9G 3D Multi-Robot TAMP & 6-Cell Smart Factory Lite-Sim," deployed live at https://tamp.mincasurong.com (2026).
  2. Lazy-DaSH (M9G-TR-2026-05 & IEEE T-RO): Seongwon Lee, James Motes, Isaac Ngui, Marco Morales, and Nancy M. Amato (UIUC Parasol Lab), "Lazy-DaSH: Lazy Approach for Hypergraph-based Multi-Robot Task and Motion Planning," IEEE Transactions on Robotics (T-RO), arXiv:2504.05552; see also M9G-TR-2026-05.
  3. Project STAMP (M9G-TR-2026-03): Minsu Chang, Ph.D., "Project STAMP: Decoupling Multimodal VLA Reasoning from Robot Execution via Deterministic Pre-Dispatch Gating and Fault Write-Back," M9G-TR-2026-03 (2026).
  4. Multi-Robot Assembly, Deformables & Self-Driving Labs (M9G-TR-2026-04): Minsu Chang, Ph.D., "Multi-Robot Manipulation in NVIDIA Isaac Sim (2025–2026): Cooperative Assembly, Deformable Handling, and Self-Driving Labs," M9G-TR-2026-04 (2026).
  5. Symbolic 3D Grounding & Software-Pipelined VLA Task Planning (M9G-TR-2026-01): Minsu Chang, Ph.D., "Symbolic 3D Grounding and Software-Pipelined VLA Task Planning: ER-2 vs. Gemini 4 Carbon," M9G-TR-2026-01 (2026).
  6. Franka Emika FR3 Kinematics & Craig Modified DH Convention: J. J. Craig, Introduction to Robotics: Mechanics and Control, 4th ed., Pearson, 2017; Franka Robotics FR3 URDF & DH Specification.
Research Provenance & Citation

This technical report (M9G-TR-2026-06) documents the client-side WebGL architecture, TypeScript 7-DoF Modified DH and 6×7 Geometric Jacobian DLS IK solver in fr3Kinematics.ts, 128 Hz swept-capsule safety gate, Lazy-DaSH hypergraph replanner, and 6-Cell Smart Factory clearance engineering deployed at https://tamp.mincasurong.com.

Minsu Chang, Ph.D. (M9G), "Engineering tamp.mincasurong.com: Zero-Install 3D Multi-Robot TAMP and a 6-Cell Smart Factory in 60 FPS WebGL," M9G Technical Research Blog (M9G-TR-2026-06), Nov. 9, 2026. Live Simulator: https://tamp.mincasurong.com | Article: https://mincasurong.github.io/blog/tamp-fast-sim/