Cam SmithAI Studio

← All models

Computational Model Studio · Model release

Neuro-symbolic Robot Manipulation (ICRA 2023)

Research hybrid that grounds a natural-language instruction in an initial depth-image scene to produce a symbolic manipulation program and pose sub-goals.

Developer: Namasivayam Kalithasan et al. (IIT Delhi) · Last verified:

What it is designed to do

Infer a high-level symbolic manipulation program and grounded pose sub-goals from a natural-language instruction and an initial depth-image scene; a separate motion planner is assumed for physical trajectories.

Primary contribution
inference planning
Input
text, image
Output
structured symbols, spatial (grounded manipulation program, object and target-pose sub-goals)

Documented uses: Author evaluation of language-guided manipulation tasks in a simulated 7-DOF robot environment.

Access and deployment

  • NSRM author checkpoint — available. Checkpoint is publicly fetchable; use it with the same-revision model code and symbolic executor. Checkpoint weight rights are not expressly documented. Artifact documentation Local hardware requirements have not been verified. The repository MIT text covers software and associated documentation but does not expressly identify checkpoint weights; weight-use and redistribution rights were not established.
Technical profile and review notes
Construction
hybrid
Representation
neural, symbolic
Normalized architecture tags
No normalized architecture tag assigned
Lifecycle
Unknown from reviewed sources

Inclusion rationale: Provides an author-named bounded hybrid with learned visual/language representations and explicit symbolic program execution.; Adds an instruction-and-scene to manipulation-program function without treating the surrounding robot runtime as the model.

Limitations

  • The public checkpoint requires same-revision model code and symbolic executor to represent the full hybrid model.
  • The author-reported experiments are simulated; physical-robot deployment is not established.
  • A separate low-level motion planner is assumed for trajectories.

Sources and related Studio records

Sources

  1. Learning Neuro-symbolic Programs for Language Guided Robot Manipulation project pageIIT Delhi research team · checked 2026-09-30
  2. Learning Neuro-symbolic Programs for Language Guided Robot ManipulationNamasivayam Kalithasan and coauthors · checked 2026-09-30
  3. NSRM source repository at pinned commitdair-iitd · checked 2026-09-30
  4. NSRM model_release.pth checkpoint at pinned commitdair-iitd · checked 2026-09-30
  5. NSRM repository license at pinned commitdair-iitd · checked 2026-09-30

Connections