Skip to content

AIC_OVERVIEW.md

Latest commit

 

History

History
158 lines (123 loc) · 6.47 KB

AIC_OVERVIEW.md

File metadata and controls

158 lines (123 loc) · 6.47 KB

AI for Industry Challenge — Competition Overview

What Is AIC?

The AI for Industry Challenge (AIC) is a robotics competition hosted by Intrinsic AI (Alphabet/Google's industrial robotics division). The task is to build an autonomous policy that can insert fiber-optic connectors (LC/SC/SFP cables) into ports on a task board, using only wrist cameras and force-torque feedback — mimicking real factory cable-routing work.

Competition Format

Timeline

  • Challenge start: March 2, 2026
  • Qualification phase: ~2.5 months (internal freeze ~May 27, outer bound June 30, 2026)
  • Full challenge end: September 2026

Environment

  • Robot: UR5e arm with a parallel-jaw gripper
  • Task board: Randomized rail positions and angles, with connector ports
  • Cables: SFP module, LC plug, SC plug
  • Simulator: Gazebo (ROS 2 Kilted) — the official evaluator
  • Cloud eval hardware: Single NVIDIA L4 (24 GB VRAM) — your policy must fit here

What a Trial Looks Like

  1. Robot starts already holding the plug
  2. Target port is guaranteed visible in at least one wrist camera
  3. Robot must approach, align, and insert — one insertion per trial
  4. Trial is scored and logged to scoring.yaml

Observation Interface (20 Hz)

Modality Detail
3x wrist RGB images left, center, right
3x CameraInfo intrinsics per camera
Joint states position, velocity, effort
Wrist wrench 6-axis F/T sensor
Controller state current impedance/mode

Scoring System

Source of truth: docs/scoring.md (NOT scoring_tests.md)

Tier Achievement Points
Tier 1 Approach — cable moved into target region up to 6
Tier 2 Alignment — plug oriented at port entrance up to 12
Tier 3 Partial insertion up to 6
Full insertion Plug fully seated 75

Penalties (score reductions)

  • Wrist force > 20N
  • Contact with off-limit zones (e.g., wrong ports, board edges)
  • Excessive trial duration

Winning means maximizing full insertions (75 pts each) while minimizing penalties.

The Competition Aims

  1. Advance physical AI — demonstrate that learned policies can handle high-precision, contact-rich manipulation in unstructured environments
  2. Sim-to-real generalization — policies trained in simulation must work on the randomized eval setup without GT poses
  3. Industrial relevance — cable routing/insertion is a real unsolved bottleneck in electronics manufacturing
  4. Multi-modal perception — teams must fuse wrist RGB + force/torque without depth sensors
  5. Safe compliance — policies must be force-aware to avoid damaging cables or hardware

How Novelty Is Explored

Why this is genuinely hard

  • No depth sensor — you only get RGB + F/T
  • Task board geometry is randomized — no fixed port positions
  • Connector insertion requires ~1mm precision
  • Cable compliance means the object deforms unpredictably
  • Force penalties punish naive stiff control

Novelty vectors teams are exploring

1. Perception

  • Keypoint detectors for plug-tip and target port localization from RGB only
  • Cable state estimation (HANDLOOM-style iterative tracers for centerline)
  • Multi-view fusion (left + center + right cameras for depth cues without a depth sensor)

2. Control

  • Variable impedance: dynamically switching stiffness/damping per task phase
    • High stiffness in free-space approach (speed)
    • Low stiffness during insertion (compliance, avoids >20N penalty)
  • Force-guided primitives: using wrist wrench to detect contact and switch phases
  • MuJoCo gain sweeps → fast Bayesian optimization of impedance parameters

3. Learning Paradigms

  • Teacher-Student Distillation (dominant approach):
    • CheatCode (GT oracle) generates perfect demos at scale
    • ACT/Diffusion policy (student) trains on oracle's RGB → action pairs
    • Student deploys with zero GT access
  • Asymmetric Actor-Critic (LUPI):
    • Critic sees full GT during RL training
    • Actor sees only cameras + joint states
    • Only Actor deployed at eval time
  • Hybrid: analytic state machine wrapper around a learned spatial policy

4. Data

  • Synthetic GT demos (CheatCode in Gazebo/Isaac)
  • Human teleop demos (LeRobot + AIC keyboard teleop)
  • Isaac Lab parallel envs for massively scaled RL (AIC-Task-v0, rsl-rl)
  • Cross-simulator consistency (MuJoCo ↔ Gazebo policy transfer)

5. Architecture

  • ACT (Action Chunking with Transformers) — predicts action sequences
  • Diffusion Policy — stochastic, handles multimodal action distributions
  • Hybrid analytic + learned (state machine wrapping neural net output)

How to Win — Fundamentally

The Core Loop

CheatCode (GT oracle, Gazebo)
  → 10,000+ perfect demo recordings (RGB images + joint velocities)
  → Train ACT/Diffusion student policy
  → State machine wrapper (approach: stiff → insertion: compliant)
  → Non-GT perception (keypoint detectors for plug-tip + port)
  → Validate ALL decisions in Gazebo (truth lane)
  → Submit policy package that fits L4 (24 GB)

The Three-Lane Development Model

Lane Simulator Purpose
Truth Gazebo / Kilted Every real decision validated here, scoring.yaml is the metric
Controller MuJoCo mirror Fast gain/impedance sweeps without Gazebo overhead
Learning Isaac Lab Parallel RL, demo recording, policy pretraining

Priority Order

  1. Tier 3 > 0 without GT — prove your perception can localize the port
  2. Full insertion > 50% — prove your insertion primitive is reliable
  3. Force penalties < 10% — prove your compliance switching works
  4. Duration minimized — speed matters for ranking when insertion rates are equal

The One-Line Strategy

Train a compliant, force-aware student policy via Teacher-Student Distillation from a GT oracle, with a state-machine wrapper that drops stiffness at contact, validated entirely in Gazebo.

Key Repos

Repo Role
intrinsic-dev/aic Official competition repo
huggingface/lerobot ACT, Diffusion Policy, teleop recording
vainaviv/handloom Cable state estimation (perception)
MarkFzp/act-plus-plus ACT++ training loops
real-stanford/diffusion_policy Diffusion Policy baseline
NVlabs/industreallib Insertion controller patterns (Franka-based donor)
RMDLO/trackdlo DLO shape tracking (RGB-D, may need adaptation)