Skip to content

v0.1.0-aht

feat(aht): add ego-AHT HAPPO wrapper + runner + env adapter (v0.1.0-aht)
@fsafaei fsafaei tagged this 09 May 20:41 
Lands the CONCERTO ego-AHT extensions on top of upstream HARL per
ADR-002 §Decisions and plan/05 §3 (CONCERTO project — see
phase0_reading_kit/plan/05-training-stack.md). Three new modules; no
existing HARL code is patched (plan/05 §2: minimise upstream merge
conflicts).

New files:

- harl/algorithms/actors/ego_aht_happo.py
  EgoAHTHAPPO subclass of HAPPO. Validates the partner is frozen at
  construction (ADR-009 §Consequences runtime backstop for the AHT
  no-joint-training contract); ego-only advantage decomposition;
  reduces to single-agent PPO update. ``# UPSTREAM-VERIFY:`` markers
  flag every spot that depends on the pinned upstream HAPPO signature
  (collect_rollout body, update body, from_config kwargs).

- harl/runners/ego_aht_runner.py
  Hydra-driven launcher. Validates the composed config through
  CONCERTO's EgoAHTConfig (Pydantic v2), then delegates to
  chamber.benchmarks.training_runner.run_training with
  EgoAHTHAPPO.from_config plugged into the trainer-factory seam from
  CONCERTO M4b-5. Recommended invocation:
      python -m harl.runners.ego_aht_runner \
          --config-path "$PWD/configs/training/ego_aht_happo" \
          --config-name mpe_cooperative_push

- harl/envs/concerto_env_adapter.py
  Thin shim from CONCERTO's Gymnasium-multi-agent env to HARL's
  runner expectations. No new state; pure pass-through with a
  ``# UPSTREAM-VERIFY:`` note about HARL's pre-/post-Gymnasium-0.26
  return-tuple convention.

- tests/test_ego_aht_happo.py
  Subclassing + frozen-partner-validation smoke tests (skipped until
  the upstream HAPPO __init__ signature is filled into
  _HAPPO_INIT_ARGS). The two non-skipped tests cover the test
  fixtures themselves so the freeze-check helper is sanity-checked.

Inheritance contract (CONCERTO ↔ this fork after v0.1.0-aht):

- concerto.training.ego_aht.train(cfg, *, env, partner, trainer_factory)
  is the algorithm-agnostic loop (CONCERTO M4b-5; concerto.* does not
  import chamber.*).
- chamber.benchmarks.training_runner.run_training(cfg, *, trainer_factory)
  is the chamber-side bridge (CONCERTO M4b-5).
- harl.algorithms.actors.ego_aht_happo.EgoAHTHAPPO.from_config is the
  fork-side trainer-factory: it satisfies CONCERTO's
  concerto.training.ego_aht.TrainerFactory Protocol structurally and
  builds an EgoAHTHAPPO from the validated EgoAHTConfig.

A follow-up CONCERTO PR will:
1. Bump pyproject.toml's harl @ git+...@<SHA> to this v0.1.0-aht SHA.
2. Add tests/integration/test_ego_aht_runner_dry.py exercising the
   full plumbing on a 100-frame dry run.
3. Replace the M4b-8 empirical-guarantee experiment's RandomEgoTrainer
   default with EgoAHTHAPPO.from_config so T4b.13 actually exercises
   the AHT loop (rather than uniform-random ego actions).

Licence: HARL is MIT-licensed; the fork inherits MIT for every file
copied from upstream. The four files added by this commit are
Apache 2.0 / MIT-compatible per their SPDX header
(SPDX-License-Identifier: Apache-2.0) — see the NOTICE update
accompanying this commit ("Modifications by CONCERTO Contributors,
licensed under Apache 2.0").
Assets 2
Loading