Re Real Time Crowd Navigation From Fi First P Princi ncipl ples - - PowerPoint PPT Presentation

Re Real Time Crowd Navigation From Fi First P Princi ncipl ples o

• f P

Proba babi bility The Theory

Pete Trautman and Karankumar Patel Honda Research Institute San Jose, CA, USA

What is Crowd Navigation?

• Unstructured: no flow rules or static obstacles
• Collect time stamped x,y data (trajectories)

True Pedestrian Movement: Humans leverage cooperation Goal Desired Path Executed Path Even with perfect future knowledge, “decoupling” prediction and navigation fails

Decoupling leads to Freezing Robot Problem

Optimality Theorems of Autonomous Crowd Navigation. CDC, 2017. [43] P. Trautman and A. Krause. Unfreezing the robot: Navigation in dense interacting crowds. In IROS, 2010.

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 mgGP # unsafe/# runs: mgIGP # unsafe/# runs: No unsafe runs 5/19 2/30 12/22 6/28 12/16 2/13 17/21 2/7 19/21 4/9 7/8 3/3 8/8 2/2 Average Crowd Density Over Duration of Run (people/m

2)

Fraction of Unsafe Runs % mgGP unsafe: 0.63492, total runs: 126 % mgIGP unsafe: 0.19444, total runs: 108

Large safety decrement at all densities

Decoupling: Suboptimal at any Density

Trautman, P.; Ma, J.; Krause, A.; and Murray, R. M. 2013. Robot navigation in dense crowds: the case for cooperation. In ICRA. Trautman, P. 2013. Robot Navigation in Dense Crowds: Statistical

Necessity of Coupled Models

u(t + 1) = f R∗(t + 1)

[f R, h]∗ = arg max

f R,h

p(h, f R | zh

T, zR T)

where p(h, f R | zh

T, zR T) = ψ(h, f R, γ)p(f R | zR T)p(h | zh T)

p(f R, h1, h2 | z)

[f R, h1, h2]∗

• What must ψ(h, f R, γ) model?
• How can ψ(h, f R, γ) model it?

The How and What of Interaction

• vercautious (FRP) robot

⇒ cost only a function of the full set of mixture statistics ⇐ ⇒ C ⇣ µR

` (t), µf i k (t),

| {z }

intents at t

w(t) |{z}

preference

, Σ(t) |{z}

flexibility

⌘

→ Provide statistically valid interaction function PIGP

¬κ

→ Provide a real time locally optimal solver

Crowd Navigation Evaluation Challenges

• Real world deployment

expensive: 6 months-year to deploy statistically valid study

• ORCA, Social Forces based

simulation non-discriminative

• Optimal policy is ”blind, straight

line”

ORCA sim collision rate, 0.05-0.25 people/m^2

ETH: 241 runs “Leave one out” evaluation:

• Remove 1 human
• Start and goal
• Compare

safety/efficiency of robot and human

Evaluation: Partial Runs

(a) (b)

Conclusion

• Provided constraints on permissible

interaction functions (for GP mixtures)

• Flexibility key to mitigating freezing

robot problem

• Provided a discriminative evaluation

scenario Next steps:

• Deploy in real world