# Enabling High-Quality Untethered Virtual Reality (MoVR)
Abari _et al.,_ USENIX NSDI '17

## Motivation/Introuction
- Streaming Virtual Reality (VR) systems
	- Multi-Gbit/s data rates from game console to wearable headset
	- 100 frames per second frame rate to make it realistic, avoid nausea
- Replace cable from console to headset with RF link
	- mmWave: only practical RF technology that supports those rates
		- > 24 GHz, 24 GHz in this paper
		- Highly directional, like light
		- **Problem:** Blocked by hand, head, obstacles, like light
	- **Challenge:** User is mobile, creates dynamic blockages
- Interest from Google, many other smaller companies

##  Design (§3)
- Overview shown in Figure 2: wall-mounted mmWave relay module (referred to as the "_mirror_" throughout)
	- AP chooses to send direct (above), or thru relay if direct path blocked
- **Quantifying blockage** (Figure 4)
	- Hand: -14 dB, head: -20 dB, body: -25 dB
	- **Status quo:** AP finds a "bounce" path off a wall or object (best case of *that* is **-15 dB**, since longer distance and bounce scattering)

### MoVR Mirror Design (§4.2)
- Overall design (Figure 6): receive antenna to amplifier to transmit antenna
	- Phased array antennas for both transmit & receive
		- Create narrow transmit, receive beams
	- No radio, just a relay
	- **Problem:** Some signal transmitted from mirror is received by mirror, resulting in **feedback loop** that saturates the amplifier, generates garbage (Fig. 6b, right)
	- So set the amplifier so that forward _gain_ minus _leakage_ is negative, so the feedback loop dies away
		- So need forward gain **less than** leakage
		- **But leakage varies** with angles of antennas (Figure 7)
		- So need **adaptive algorithm** to set the forward amplifier gain
	- **Challenge:** No radio to measure signal levels on the mirror
	- **MoVR approach:** Monitor current draw of the amp, increase gain continuously until **high current draw** indicates saturation, then back off the gain.  Repeat continuously.

### Alignment and Tracking (§§5.1, 5.2)
- Very sensitive to alignment (-20 dB for 10 degrees misalignment as shown in Figure 9), so need precise alignment
	- *For Direct Mode:* Need AP to headset alignment
	- *For Relay Mode:* Need AP to mirror, **and** mirror to headset alignment

- **b) AP to headset** alignment:
	- Just leverage VR laser tracker: gives location _and_ orientation
- **a) AP to mirror** alignment:
	- Challenge: The mirror can't receive, just relay
	- AP transmits on some frequency, mirror reflects **back to the AP** on a neighboring frequency, AP measures power of the reflection
		- Frequency shift overcomes problems with listening and transmitting on the same frequency
		- **Need two angles:** Angle at AP $\theta_1$, angle at mirror $\theta_2$
			- Exhaustive two-dimensional search in those two angles to maximize power
- **c) Mirror to headset** alignment:
	- Leverage AP to headset alignment from VR laser tracker, from step (a) above
	- Then, translate coordinates from AP to mirror: **have angle** from step (b) above, **need distance** as well to translate coordinates
	- Approach: **Triangulate mirror location** as shown in **Figure 10**
		- $\phi_{AP}$ is $\theta_1$ from previous step
		- $\phi_H$ is found by a similar two-dimensional exhaustive search with AP sending to mirror, searching over mirror and headset angles


## Evaluation (§7)
- HTC VIVE VR system hardware with custom mirror hardware in a 5 meter by 5 meter room
- LoS blocked 20 times over 5 minutes while playing a game (**Fig. 12** shows blockage durations, 245 milliseconds median, _i.e._ 25 frames, so would cause a noticeable glitch)
- Mirror Performance (no blockage = 0 dB reference level in Fig. 13)
	- Compare no blockage versus blockage without MoVR
		- Fig. 13: -27 dB loss from bouncing off a wall
	- Compare no blockage versus blockage **with** MoVR
		- Fig. 13: Improvement in SNR from amplification, shorter paths
	- Why sometimes 3 dB worse?  No explanation, possibly amplifier saturation, beam alignment problems
- Beam Alignment and Tracking Performance (§7.3)
	- Two degree accuracy for AP to mirror beam alignment (Fig. 14), ground truth coming from laser distance measurement tool
	- Whole system (Fig. 15) versus exhaustive search in **all three angles:** MoVR 4 dB worse sometimes about the same number of times to possibly explain Fig. 13
- Beam Alignment Latency
	- **Important:** needs to be fast to track -- claimed 0.9 $\mu$s to align, 1.7 $\mu$s to reconfigure phase shifters, compared to 1 ms delay in the VR tracker so MoVR is much faster
- System Performance (§7.4)
	- Compare no mirror (bounce off walls), fixed gain mirror, MoVR with **LoS path blocked by hand**
	- Where's the AP in Fig. 16?
	- Better with MoVR's adaptive gain control