Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Following the Human Thread in Social Navigation

Authors: Luca Scofano, Alessio Sampieri, Tommaso Campari, Valentino Sacco, Indro Spinelli, Lamberto Ballan, Fabio Galasso

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Tested on the novel Habitat 3.0 platform, SDA sets a novel state-of-the-art (Sot A) performance in finding and following humans. Out of extensive benchmarking, SDA outperforms the approach proposed in Habitat 3.0 (Puig et al., 2024) and a second adapted best-performing method (Cancelli et al., 2023) from Habitat 2.0 (Szot et al., 2021). We conduct a thorough experimental evaluation of the core contribution of this work learning to infer social dynamics from (privileged) information about the person. Our ablative studies reveal that human trajectories are not only strong input information for the robot control policy but also provide better supervision for inferring the social dynamics latent to the same policy.
Researcher Affiliation Collaboration 1 Sapienza University of Rome 2 Fondazione Bruno Kessler 3 University of Padova 4 Ital AI
Pseudocode No The paper describes the methodology in text and with diagrams (Figure 1 and Figure 2), but it does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code can be found at https://github.com/L-Scofano/SDA.
Open Datasets Yes Tested on the novel Habitat 3.0 platform... This constraint has been effectively addressed with Habitat 3.0 (Puig et al., 2024), the simulator used for this research.
Dataset Splits No The paper mentions training for "250 million steps across 24 environments" for Stage 1 and "5 million steps across the same environments" for Stage 2, but it does not specify explicit training, validation, or test dataset splits in terms of percentages or sample counts for the environments or data within them.
Hardware Specification Yes Both stages utilize 4 A100 GPUs for efficient computation.
Software Dependencies No The paper mentions DD-PPO (Wijmans et al., 2019b) and ResNet (He et al., 2016) as key components but does not provide specific version numbers for these or other software libraries or frameworks used in the implementation.
Experiment Setup Yes In Stage 1, we utilize DD-PPO (Wijmans et al., 2019b) for 250 million steps across 24 environments, following the training protocol presented in (Puig et al., 2024). Furthemore, each time step is approximately 0.04 seconds, so we consider a trajectory of 0.8 seconds. The trajectory encoder (ยต) is implemented as Multilayer Perceptrons (MLPs). The Adapter module, parametrized by an MLP ฯˆ that takes as input the recent history of the robot s states xt N:t 1 and actions at N:t 1 to generate a new latent vector ห†zt. The output zt has a dimensionality of 128.