Characterizing and Improving the Resilience of Accelerators in Autonomous Robots

Deval Shah, Zi Yu Xue, Karthik Pattabiraman, Tor M. Aamodt

TL;DR

Autonomous robots rely on specialized hardware accelerators for motion planning, but these accelerators are vulnerable to soft errors that can cause dangerous collisions. This paper introduces a novel metric, the Collision Exposure Factor (CEF), to identify the most critical bits in these accelerators. By using CEF to guide selective error mitigation and a new fault injection technique, the authors achieve significant improvements in resilience with minimal overhead, making autonomous robots safer and more reliable.

What Problem Does This Paper Address?

Autonomous robots are increasingly deployed in safety-critical applications like self-driving cars and home assistance. A core component of these robots is motion planning, which enables them to navigate and avoid obstacles. To meet the demanding real-time performance requirements, specialized hardware accelerators (MPAs) are used. However, these MPAs are susceptible to "soft errors" – transient faults caused by radiation or other environmental factors. A single soft error in an MPA could lead to a miscalculation in collision detection, causing the robot to collide with an object, posing a significant safety risk. Traditional methods to protect against soft errors (like ECC or redundancy) are often too costly in terms of power, area, and performance for consumer-grade robotics. Therefore, there's a critical need for efficient and targeted methods to ensure the reliability of these MPAs without prohibitive overheads.

Methods

The paper introduces two main innovations:

  • Collision Exposure Factor (CEF): This novel metric quantifies the vulnerability of individual bits within a motion planning accelerator's collision detection module (CDM). CEF is based on the idea that a bit flip is more likely to cause a safety violation (a collision) if it exposes a larger area of the robot's potential path to obstacles. It's calculated by considering the "critical space" of a bit – the part of the swept space that would be excluded if that bit flips – and measuring its exposed surface area. Crucially, CEF is independent of the specific environment or obstacle positions, making it a general metric for a given robot and MPA.
  • CEF-aware Fault Injection (FI) Methodology: To efficiently characterize the impact of soft errors and validate CEF, the authors propose a two-phase FI approach:
  • Phase 1 (CEF Measurement): This phase uses microarchitectural simulation to inject single-bit errors and calculate the CEF for every bit in the MPA's storage. This is done without considering specific environments, making it fast.
  • Phase 2 (SDC-C Probability Measurement): Leveraging the CEF values from Phase 1, this phase groups bits by CEF and samples from these groups to estimate the probability of a soft error causing a collision (SDC-C). This significantly reduces the number of FI runs compared to exhaustive methods.

These methods are used to guide CEF-aware selective error mitigation, where only the most critical bits (identified by high CEF) are protected using techniques like ECC or latch hardening.

Datasets

The paper does not explicitly mention training datasets. For evaluation, they used:

  • Four different Collision Detection Modules (CDMs): Labeled A1, A2, A3, and A4, representing various architectural approaches (voxel-based, box-based, octree-based).
  • Three robotic arms: Jaco2, AL5D, and Puma761, simulated using the Klampt software.
  • Four benchmark environments (D1-D4): Characterized by varying obstacle densities, used to study the impact of environment on SDC-C probability.

Key Insights and Ideas

  • CEF is a strong predictor of SDC-C: The paper empirically demonstrates a strong positive correlation between a bit's CEF and its probability of causing a collision. This validates CEF as a reliable metric for assessing vulnerability.
  • Selective protection is highly effective: Protecting only the most vulnerable bits (identified by CEF) provides a significantly better FIT rate reduction for the same amount of protected memory compared to uniform, bit-position, or access-frequency-based protection.
  • CEF-aware FI drastically reduces testing time: The proposed two-phase CEF-aware FI method achieves a speedup of up to 23,000x over exhaustive FI, making detailed reliability analysis feasible.
  • Geometric representation impacts vulnerability: Different geometric representations (voxel, box, octree) for swept space lead to distinct CEF distributions, influencing the effectiveness of error mitigation strategies. Octree-based representations, for instance, tend to have higher CEF values for certain bits due to their hierarchical nature.

Key Results

  • FIT Rate Reduction: CEF-aware selective error mitigation achieved an average FIT rate reduction of 12.3x, 9.6x, and 4.2x compared to uniform, bit-position, and access-frequency-aware methods, respectively, for the same protected memory.
  • FI Speedup: The CEF-aware FI approach achieved an average speedup of 23,000x over exhaustive FI with a low error margin (2.5%).
  • Area/Power Savings: For high safety integrity levels (SIL 3 and SIL 4), CEF-aware mitigation achieved significant area and power savings compared to full protection, with overheads as low as 3% for A1 and 1.8% for A2.
  • CEF Distribution: The distribution of CEF values across bits is highly asymmetric, with a large number of bits having a CEF of 0, indicating they require minimal or no protection.

Limitations and Future Work

  • Single-bit fault model: The current work focuses on single-bit flips. The authors suggest extending the CEF metric and FI methods to handle multi-bit fault models.
  • Analytical CEF model: While the current CEF calculation relies on FI, the authors propose developing an analytical model to compute critical space and CEF without the need for extensive simulations.
  • Generalizability: While the CEF concept is applicable to other robotics accelerators processing spatial information, the specific implementation of Algorithm 1 would need modification for different geometric representations (e.g., triangle meshes).
  • Resilience-aware co-design: The paper identifies architectural parameters affecting resilience and suggests future work on "resilience-aware" algorithm-hardware co-design.

Summary updated

Abstract

Motion planning is a computationally intensive and well-studied problem in autonomous robots. However, motion planning hardware accelerators (MPA) must be soft-error resilient for deployment in safety-critical applications, and blanket application of traditional mitigation techniques is ill-suited due to cost, power, and performance overheads. We propose Collision Exposure Factor (CEF), a novel metric to assess the failure vulnerability of circuits processing spatial relationships, including motion planning. CEF is based on the insight that the safety violation probability increases with the surface area of the physical space exposed by a bit-flip. We evaluate CEF on four MPAs. We demonstrate empirically that CEF is correlated with safety violation probability, and that CEF-aware selective error mitigation provides 12.3x, 9.6x, and 4.2x lower Failures-In-Time (FIT) rate on average for the same amount of protected memory compared to uniform, bit-position, and access-frequency-aware selection of critical data. Furthermore, we show how to employ CEF to enable fault characterization using 23,000x fewer fault injection (FI) experiments than exhaustive FI, and evaluate our FI approach on different robots and MPAs. We demonstrate that CEF-aware FI can provide insights on vulnerable bits in an MPA while taking the same amount of time as uniform statistical FI. Finally, we use the CEF to formulate guidelines for designing soft-error resilient MPAs.
BibTeX
@misc{shah2021characterizing,
  author = {Deval Shah and Zi Yu Xue and Karthik Pattabiraman and Tor M. Aamodt},
  title = {{Characterizing and Improving the Resilience of Accelerators in Autonomous Robots}},
  year = {2021},
  eprint = {2110.08906},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2110.08906},
}

Loading related content...

Figure