M A HAFIZ - Results & Impact

Proof that my systems work (not just “cool demos”)

This page highlights measurable outcomes across my robotics work: simulation fidelity, stable control, task success, and repeatable evaluation — backed by artifacts (metrics, reports, and videos).

Summary: A candidate controller can look “fine” (even 100% success), yet still be unsafe to ship. My platform detects such regressions automatically and outputs a clear SHIP / BLOCK decision.

See Projects See Architecture Download CV Contact

100%

Task Success Rate (Demo run)

Measured over 5 episodes in the Robot Eval Platform demo.

20.8 ms

Mean Control Latency (Candidate)

Regression detected: +3.6 ms (+21%) vs baseline 17.2 ms.

BLOCK

Release Gate Decision (Do not ship)

Candidate automatically blocked due to violating the latency regression rule.

Robot Eval Platform — Regression Gating Outcomes

Automated rollouts into metrics + videos + reports for every run.
Baseline vs candidate comparison with a clear SHIP / BLOCK decision.
Rules show exactly why a candidate is blocked (traceable + auditable).

FastAPIReactPostgreSQL MinIO/S3CI-style gating

Example: Candidate achieved 100% success but was BLOCKED because control latency regressed by +21% vs baseline.

Gate rules used in the demo

Success rate ≥ baseline
Mean control latency ≤ baseline
Safety violations = 0

This is the “decision layer” that prevents silent regressions from reaching real robots.

Evidence (Screenshots + Demo Video)

Demo video (rollout + artifacts)

This demo shows: run execution → rollout video → metrics artifacts → gating decision (SHIP/BLOCK).

Key screenshots

Overview: Candidate blocked despite 100% success due to latency regression.

Compare: Baseline 17.2 ms vs candidate 20.8 ms (+3.6 ms).

Runs: History view showing SHIP/BLOCK decisions across runs.

Gate details: Rule-level traceability (exact failure reason shown).

Run details: Rollout preview stored as an artifact for auditability.

Settings: Visible evaluation policy (metrics enforced + strategy).

Franka FR3 — Gesture-Driven UX + Sim2Real Validation (Free-space + Contact-rich)

3.06 mm

Best Final Reach Error (Task 1)

Participant P02, Trial 5 (Reach task) achieved 3.06 mm final positioning error.

≈99%

Learning Improvement (Reach task)

Large initial error reduced to single-digit mm (e.g., P02: 248 mm → 3 mm in 5 trials).

Up to 38.9 N

Contact Force Peak (Task 2)

Threaded insertion/screwing produced controlled forces (P01 peak 38.9 N, then stabilizing).

Hand-gesture-driven UX (EMG/IMU) for axis selection + action execution via lock/unlock state machine.
Cartesian motion + orientation control executed through Jacobian/IK-based control.
Validated on real FR3 across both: Task 1 (free-space reach) and Task 2 (contact-rich screw depth).
Objective metrics recorded from the FR3: end-effector pose error, force/torque behavior, and rotation alignment.

FR3Gesture UXEMG/IMU Cartesian ControlImpedanceContact-rich

Key point: This is not “gesture → direct joint motion”. It is a complete HRI control stack with a UX state machine and measurable performance outcomes.

EMG / IMU
  ↓
Feature Extraction + Classification
  ↓
UX State Machine (scroll / lock / unlock)
  ↓
Cartesian Command (axis + direction)
  ↓
Jacobian/IK Controller (+ safety limits)
  ↓
MuJoCo Simulation  →  Real FR3 Deployment

Objective results snapshot (from user study)

Participant	Task 1 Reach Error (Trial 1 → Trial 5)	Improvement	Task 2 Force Range (N)	Task 2 Rotation Error
P01	322.27 mm → 7.05 mm	~97.8%	4.46 – 38.94 N	≤ 0.32° (then near 0°)
P02	248.40 mm → 3.06 mm	~98.8%	1.61 – 26.64 N	0 – 0.15°
P03	34.16 mm → 4.91 mm	~85.6%	15.11 – 20.86 N	0 – 0.07°

For Task 2 (screwing), force/torque + alignment are the more informative indicators than positional error alone.

Evidence (Gesture UX + MuJoCo Simulation + Real FR3)

MuJoCo simulation demo

Demonstrates gesture-driven UX → Cartesian command execution in simulation.

Real FR3 deployment demo

Demonstrates the same UX/control workflow deployed on real hardware (FR3).

Screenshots (UX + plots)

Gesture UX: Menu / axis selection with lock–unlock control.

Real setup: FR3 experiment setup used for the user study.

Task 1: Reach error decreases across rounds (learning effect).

Task 2: Force/torque behavior during threaded contact (screw depth).

What I optimize for

Measurable progress (metrics-first, reproducible experiments)
Stability & safety (limits, smooth control, robust behavior)
Deployment readiness (CI-style checks and clear failure modes)
Fast iteration (tight loop: simulate → evaluate → improve)