Crowd-Sourcing: Mars Escape – Overview

The development of hand-crafted models for action and dialog generation for a social robot is a time consuming process that yields solutions for only a relatively narrow range of interactions envisioned by the programmers. Further, it is important that such models be compatible with human behavior.We are exploring a data-driven approach for interactive behavior generation that leverages on-line games as a means for collecting large-scale data corpora for human-robot interaction research. We are developing methods, building on the work of Orkin and Roy’s Restaurant Game, to learn context-relevant action and dialog models from this crowdsourced behavior to support human-robot collaboration.

We have developed a 2-player on-line game called MARS ESCAPE to support this research. In the game, you can take the role of an astronaut or a robot on Mars who must complete their mission before oxygen supplies run out.  Data collected from the game will be used to generate autonomous behaviors for the MDS robot Nexi and to study human-robot interaction by recreating the game environment in real life at the Boston Museum of Science in 2011.

These learned models of typical and atypical human behavior are being applied to enable a virtual robot to collaborate with a new human player to accomplish collaborative tasks in the context of the MARS ESCAPE game. We are also investigating how such models can be applied as a kind of “episodic memory” to enable a physical robot to perform a similar task with people in the real world.

This project is in collaboration with Prof. Sonia Chernova at WPI.

This project is supported by grants from Microsoft and the Office of Naval Research.

Study Details

The Mars Escape Game was developed at the MIT Media Lab by Jeff Orkin and designed by Sonia Chernova and Elisabeth Morant to collect a large amount of data. Two players, an astronaut and a robot, must collaborate and collect a number of items in a timed virtual game. Data was collected from hundreds of online dyads resulting in hundreds of high-level first-person policy traces. These were used to train a robot to act in a collaborative fashion in a real-world environment.

Hundreds of task demonstrations were used to train a robotic policy to interact collaboratively with a human participant. Policies were trained from state-action-time triplets using just the actions taken by the participant in the robot’s role. These algorithms were used as high level policies in a human-robot interaction task with everyday users.

The task was designed to force collaboration. Five items were selected to be collected by the participants that require sometimes simultaneous actions to be taken. Some items could only be retrieved by the participant and some could only be retrieved by the robot, forcing collaborative effort to be successful in the overall task. The items are as specified:

  • Toxic Canister: Toxic to the human, only the robot could retrieve it.
  • Alien: Either the participant or the robot can retrieve this item. Collaboration is optional.
  • Journal: High on the bookshelf, only the participant can reach it.
  • Computer Chip: This item was hidden in a lock-box that can only be opened by having both the human and robot stand in the same place.
  • Lifeform: Any one of 100 small boxes could contain a small creature. The participant can go through them one by one but it would be more efficient to ask the robot to xray all of them at once and report back the secret location of the lifeform.


  • Cynthia Breazeal, Nick DePalma, Jeff Orkin, Sonia Chernova, Malte Jung. (2013). Crowdsourcing human-robot interaction: New methods and system evaluation in a public environment. Journal of Human-Robot Interaction, 2(1), 82–111.
  • Sonia Chernova, Nick DePalma, Elisabeth Morant, Cynthia Breazeal. (2011). Crowdsourcing human-robot interaction: Application from virtual to physical worlds. In RO-MAN, 2011 IEEE (pp. 21–26). IEEE.
  • Malte Jung, Nick DePalma, Sonia Chernova, Pamela Hinds, Cynthia Breazeal. (2012). Human-Robot Collaboration: Bids and Bytes. In Proceedings of the 2012 HRI Workshop on Human-Agent-Robot Teamwork. ACM.
  • Sonia Chernova, Jeff Orkin and Cynthia Breazeal. Crowdsourcing HRI through Online Multi-Player Games. In the AAAI Fall Symposium on Dialog with Robots, 2010.
  • Sonia Chernova, Nick DePalma, Cynthia Breazeal. Crowdsourcing Real World Human-Robot Dialog and Teamwork through Online Multiplayer Games. AI Magazine, Vol 32, No 4, 2011.