analiza aiqbench
This study investigates geometric conditioning within a 0.8B hybrid language model designed for robotic manipulation, finding no consistent performance advantage from aligning training-time geometry compared to shuffled or absent conditions. The research highlights a critical gap between coordinate invariance and true physical-layout generalization, demonstrating that visual policies remain fragile even with minor object displacements. These findings offer essential diagnostics for researchers developing embodied AI systems, emphasizing the difficulty of achieving robustness in real-world spatial tasks.
Geometry Conditioning in an Embodied SLM: Training Controls and Robustness Diagnostics in a 0.8B Hybrid Model
arXiv:2609.09213v1 Announce Type: cross Abstract: We study how physical-state inputs affect a 0.8B hybrid language model adapted for manipulation with 6.2M trainable parameters. Six conditions are trained on three LIBERO-Spatial tasks and evaluated over three seeds and 540 held-out rollouts. Conditioning recurrent decay gates on geometric increments yields 28.9% success, compared with 36.7% when those increments are shuffled during training and 24.4% without explicit object/goal geometry. Both geometry policies receive correct inputs at evaluation. A token adapter using the same increments scores 27.8%; differences vary across seeds and remain inconclusive. Token-clock conditioning scores 11.1%, including one seed that fails to converge. In separate robustness tests, a state-only relative-coordinate policy retains 7/10 success under frame relabeling, whereas all four tested visual policies fall to at most 3/20 after a 5 cm object displacement. These results show no reliable advantage from training-time geometric alignment under this recipe and illustrate the gap between coordinate invariance and physical-layout generalization. Episode records, seed-level analyses, and figure-generation code accompany the paper.
Title:Geometry Conditioning in an Embodied SLM: Training Controls and Robustness Diagnostics in a 0.8B Hybrid Model
Abstract:We study how physical-state inputs affect a 0.8B hybrid language model adapted for manipulation with 6.2M trainable parameters. Six conditions are trained on three LIBERO-Spatial tasks and evaluated over three seeds and 540 held-out rollouts. Conditioning recurrent decay gates on geometric increments yields 28.9% success, compared with 36.7% when those increments are shuffled during training and 24.4% without explicit object/goal geometry. Both geometry policies receive correct inputs at evaluation. A token adapter using the same increments scores 27.8%; differences vary across seeds and remain inconclusive. Token-clock conditioning scores 11.1%, including one seed that fails to converge. In separate robustness tests, a state-only relative-coordinate policy retains 7/10 success under frame relabeling, whereas all four tested visual policies fall to at most 3/20 after a 5 cm object displacement. These results show no reliable advantage from training-time geometric alignment under this recipe and illustrate the gap between coordinate invariance and physical-layout generalization. Episode records, seed-level analyses, and figure-generation code accompany the paper.
7 pages, 3 figures. Includes ancillary data and analysis code
Robotics (cs.RO); Artificial Intelligence (cs.AI)
arXiv:2609.09213 [cs.RO]
(or arXiv:2609.09213v1 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2609.09213 Focus to learn more arXiv-issued DOI via DataCite
supp/analysis.json
supp/figures/architecture.pdf
supp/figures/architecture.tex.txt
supp/preregistered_ledger_ANONYMIZED.md
supp/records/frame_relabeling/mtS_b0state_rot0.json.episodes.jsonl
supp/records/frame_relabeling/mtS_b0state_rot45.json.episodes.jsonl
supp/records/frame_relabeling/mtS_b0state_rot90.json.episodes.jsonl
supp/records/frame_relabeling/mtS_c1state_rot0.json.episodes.jsonl
supp/records/frame_relabeling/mtS_c1state_rot45.json.episodes.jsonl
supp/records/frame_relabeling/mtS_c1state_rot90.json.episodes.jsonl
supp/records/frame_relabeling/mtS_f1state_rot0.json.episodes.jsonl
supp/records/frame_relabeling/mtS_f1state_rot45.json.episodes.jsonl
supp/records/frame_relabeling/mtS_f1state_rot90.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t0_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t0_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t0_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t4_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t4_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_c0_host_s0_t4_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t0_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t0_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t0_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t4_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t4_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_c2_relcoord_concat_s0_t4_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t0_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t0_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t0_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t4_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t4_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_f0_geometry_decay_s0_t4_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t0_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t0_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t0_cm5.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t4_cm0.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t4_cm10.json.episodes.jsonl
supp/records/layout_shift/layE_f0_shuffled_geometry_s0_t4_cm5.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_b1_clock_decay_s2_t8.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_c0_host_s2_t8.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_c1_coord_mlp_s2_t8.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_c2_relcoord_concat_s2_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_geometry_decay_s2_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s0_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s0_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s0_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s1_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s1_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s1_t8.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s2_t0.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s2_t4.json.episodes.jsonl
supp/records/main_table/m3E_f0_shuffled_geometry_s2_t8.json.episodes.jsonl
supp/records/split_manifest.json
supp/records/training_summaries.json
supp/scripts/analyze.py
supp/scripts/paired_stats.py
(92 additional files not shown) You must enabled JavaScript to view entire file list.
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 389 cuvinte
Subiecte conexe
News
Quantum Computing Report · 7 iun. 2026
How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents
NVIDIA · 26 aug. 2026
Hybrid Residual Reinforcement Learning for Contact-Rich Robotic Book Insertion
arXiv · 18 sept. 2026
Empirical Characterization of Learning Geometry in Hybrid Quantum Forecasting Models
arXiv · 21 aug. 2026